Mercurial > hg > CbC > CbC_gcc
annotate gcc/tree-ssa-threadupdate.c @ 127:4c56639505ff
fix function.c and add CbC-example Makefile
| author | mir3636 |
|---|---|
| date | Wed, 11 Apr 2018 18:46:58 +0900 |
| parents | 04ced10e8804 |
| children | 84e7813d76e9 |
| rev | line source |
|---|---|
| 0 | 1 /* Thread edges through blocks and update the control flow and SSA graphs. |
| 111 | 2 Copyright (C) 2004-2017 Free Software Foundation, Inc. |
| 0 | 3 |
| 4 This file is part of GCC. | |
| 5 | |
| 6 GCC is free software; you can redistribute it and/or modify | |
| 7 it under the terms of the GNU General Public License as published by | |
| 8 the Free Software Foundation; either version 3, or (at your option) | |
| 9 any later version. | |
| 10 | |
| 11 GCC is distributed in the hope that it will be useful, | |
| 12 but WITHOUT ANY WARRANTY; without even the implied warranty of | |
| 13 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
| 14 GNU General Public License for more details. | |
| 15 | |
| 16 You should have received a copy of the GNU General Public License | |
| 17 along with GCC; see the file COPYING3. If not see | |
| 18 <http://www.gnu.org/licenses/>. */ | |
| 19 | |
| 20 #include "config.h" | |
| 21 #include "system.h" | |
| 22 #include "coretypes.h" | |
| 111 | 23 #include "backend.h" |
| 0 | 24 #include "tree.h" |
| 111 | 25 #include "gimple.h" |
| 26 #include "cfghooks.h" | |
| 0 | 27 #include "tree-pass.h" |
| 111 | 28 #include "ssa.h" |
| 29 #include "fold-const.h" | |
| 30 #include "cfganal.h" | |
| 31 #include "gimple-iterator.h" | |
| 32 #include "tree-ssa.h" | |
| 33 #include "tree-ssa-threadupdate.h" | |
| 0 | 34 #include "cfgloop.h" |
| 111 | 35 #include "dbgcnt.h" |
| 36 #include "tree-cfg.h" | |
| 37 #include "tree-vectorizer.h" | |
| 0 | 38 |
| 39 /* Given a block B, update the CFG and SSA graph to reflect redirecting | |
| 40 one or more in-edges to B to instead reach the destination of an | |
| 41 out-edge from B while preserving any side effects in B. | |
| 42 | |
| 43 i.e., given A->B and B->C, change A->B to be A->C yet still preserve the | |
| 44 side effects of executing B. | |
| 45 | |
| 46 1. Make a copy of B (including its outgoing edges and statements). Call | |
| 47 the copy B'. Note B' has no incoming edges or PHIs at this time. | |
| 48 | |
| 49 2. Remove the control statement at the end of B' and all outgoing edges | |
| 50 except B'->C. | |
| 51 | |
| 52 3. Add a new argument to each PHI in C with the same value as the existing | |
| 53 argument associated with edge B->C. Associate the new PHI arguments | |
| 54 with the edge B'->C. | |
| 55 | |
| 56 4. For each PHI in B, find or create a PHI in B' with an identical | |
| 57 PHI_RESULT. Add an argument to the PHI in B' which has the same | |
| 58 value as the PHI in B associated with the edge A->B. Associate | |
| 59 the new argument in the PHI in B' with the edge A->B. | |
| 60 | |
| 61 5. Change the edge A->B to A->B'. | |
| 62 | |
| 63 5a. This automatically deletes any PHI arguments associated with the | |
| 64 edge A->B in B. | |
| 65 | |
| 66 5b. This automatically associates each new argument added in step 4 | |
| 67 with the edge A->B'. | |
| 68 | |
| 69 6. Repeat for other incoming edges into B. | |
| 70 | |
| 71 7. Put the duplicated resources in B and all the B' blocks into SSA form. | |
| 72 | |
| 73 Note that block duplication can be minimized by first collecting the | |
| 74 set of unique destination blocks that the incoming edges should | |
| 111 | 75 be threaded to. |
| 76 | |
| 77 We reduce the number of edges and statements we create by not copying all | |
| 78 the outgoing edges and the control statement in step #1. We instead create | |
| 79 a template block without the outgoing edges and duplicate the template. | |
| 80 | |
| 81 Another case this code handles is threading through a "joiner" block. In | |
| 82 this case, we do not know the destination of the joiner block, but one | |
| 83 of the outgoing edges from the joiner block leads to a threadable path. This | |
| 84 case largely works as outlined above, except the duplicate of the joiner | |
| 85 block still contains a full set of outgoing edges and its control statement. | |
| 86 We just redirect one of its outgoing edges to our jump threading path. */ | |
| 0 | 87 |
| 88 | |
| 89 /* Steps #5 and #6 of the above algorithm are best implemented by walking | |
| 90 all the incoming edges which thread to the same destination edge at | |
| 91 the same time. That avoids lots of table lookups to get information | |
| 92 for the destination edge. | |
| 93 | |
| 94 To realize that implementation we create a list of incoming edges | |
| 95 which thread to the same outgoing edge. Thus to implement steps | |
| 96 #5 and #6 we traverse our hash table of outgoing edge information. | |
| 97 For each entry we walk the list of incoming edges which thread to | |
| 98 the current outgoing edge. */ | |
| 99 | |
| 100 struct el | |
| 101 { | |
| 102 edge e; | |
| 103 struct el *next; | |
| 104 }; | |
| 105 | |
| 106 /* Main data structure recording information regarding B's duplicate | |
| 107 blocks. */ | |
| 108 | |
| 109 /* We need to efficiently record the unique thread destinations of this | |
| 110 block and specific information associated with those destinations. We | |
| 111 may have many incoming edges threaded to the same outgoing edge. This | |
| 112 can be naturally implemented with a hash table. */ | |
| 113 | |
| 111 | 114 struct redirection_data : free_ptr_hash<redirection_data> |
| 115 { | |
| 116 /* We support wiring up two block duplicates in a jump threading path. | |
| 117 | |
| 118 One is a normal block copy where we remove the control statement | |
| 119 and wire up its single remaining outgoing edge to the thread path. | |
| 120 | |
| 121 The other is a joiner block where we leave the control statement | |
| 122 in place, but wire one of the outgoing edges to a thread path. | |
| 123 | |
| 124 In theory we could have multiple block duplicates in a jump | |
| 125 threading path, but I haven't tried that. | |
| 126 | |
| 127 The duplicate blocks appear in this array in the same order in | |
| 128 which they appear in the jump thread path. */ | |
| 129 basic_block dup_blocks[2]; | |
| 130 | |
| 131 /* The jump threading path. */ | |
| 132 vec<jump_thread_edge *> *path; | |
| 133 | |
| 134 /* A list of incoming edges which we want to thread to the | |
| 135 same path. */ | |
| 136 struct el *incoming_edges; | |
| 137 | |
| 138 /* hash_table support. */ | |
| 139 static inline hashval_t hash (const redirection_data *); | |
| 140 static inline int equal (const redirection_data *, const redirection_data *); | |
| 141 }; | |
| 142 | |
| 143 /* Dump a jump threading path, including annotations about each | |
| 144 edge in the path. */ | |
| 145 | |
| 146 static void | |
| 147 dump_jump_thread_path (FILE *dump_file, vec<jump_thread_edge *> path, | |
| 148 bool registering) | |
| 0 | 149 { |
| 111 | 150 fprintf (dump_file, |
| 151 " %s%s jump thread: (%d, %d) incoming edge; ", | |
| 152 (registering ? "Registering" : "Cancelling"), | |
| 153 (path[0]->type == EDGE_FSM_THREAD ? " FSM": ""), | |
| 154 path[0]->e->src->index, path[0]->e->dest->index); | |
| 155 | |
| 156 for (unsigned int i = 1; i < path.length (); i++) | |
| 157 { | |
| 158 /* We can get paths with a NULL edge when the final destination | |
| 159 of a jump thread turns out to be a constant address. We dump | |
| 160 those paths when debugging, so we have to be prepared for that | |
| 161 possibility here. */ | |
| 162 if (path[i]->e == NULL) | |
| 163 continue; | |
| 164 | |
| 165 if (path[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
| 166 fprintf (dump_file, " (%d, %d) joiner; ", | |
| 167 path[i]->e->src->index, path[i]->e->dest->index); | |
| 168 if (path[i]->type == EDGE_COPY_SRC_BLOCK) | |
| 169 fprintf (dump_file, " (%d, %d) normal;", | |
| 170 path[i]->e->src->index, path[i]->e->dest->index); | |
| 171 if (path[i]->type == EDGE_NO_COPY_SRC_BLOCK) | |
| 172 fprintf (dump_file, " (%d, %d) nocopy;", | |
| 173 path[i]->e->src->index, path[i]->e->dest->index); | |
| 174 if (path[0]->type == EDGE_FSM_THREAD) | |
| 175 fprintf (dump_file, " (%d, %d) ", | |
| 176 path[i]->e->src->index, path[i]->e->dest->index); | |
| 177 } | |
| 178 fputc ('\n', dump_file); | |
| 179 } | |
| 180 | |
| 181 /* Simple hashing function. For any given incoming edge E, we're going | |
| 182 to be most concerned with the final destination of its jump thread | |
| 183 path. So hash on the block index of the final edge in the path. */ | |
| 184 | |
| 185 inline hashval_t | |
| 186 redirection_data::hash (const redirection_data *p) | |
| 187 { | |
| 188 vec<jump_thread_edge *> *path = p->path; | |
| 189 return path->last ()->e->dest->index; | |
| 190 } | |
| 191 | |
| 192 /* Given two hash table entries, return true if they have the same | |
| 193 jump threading path. */ | |
| 194 inline int | |
| 195 redirection_data::equal (const redirection_data *p1, const redirection_data *p2) | |
| 196 { | |
| 197 vec<jump_thread_edge *> *path1 = p1->path; | |
| 198 vec<jump_thread_edge *> *path2 = p2->path; | |
| 199 | |
| 200 if (path1->length () != path2->length ()) | |
| 201 return false; | |
| 202 | |
| 203 for (unsigned int i = 1; i < path1->length (); i++) | |
| 204 { | |
| 205 if ((*path1)[i]->type != (*path2)[i]->type | |
| 206 || (*path1)[i]->e != (*path2)[i]->e) | |
| 207 return false; | |
| 208 } | |
| 209 | |
| 210 return true; | |
| 211 } | |
| 212 | |
| 213 /* Rather than search all the edges in jump thread paths each time | |
| 214 DOM is able to simply if control statement, we build a hash table | |
| 215 with the deleted edges. We only care about the address of the edge, | |
| 216 not its contents. */ | |
| 217 struct removed_edges : nofree_ptr_hash<edge_def> | |
| 218 { | |
| 219 static hashval_t hash (edge e) { return htab_hash_pointer (e); } | |
| 220 static bool equal (edge e1, edge e2) { return e1 == e2; } | |
| 0 | 221 }; |
| 222 | |
| 111 | 223 static hash_table<removed_edges> *removed_edges; |
| 0 | 224 |
| 225 /* Data structure of information to pass to hash table traversal routines. */ | |
| 111 | 226 struct ssa_local_info_t |
| 0 | 227 { |
| 228 /* The current block we are working on. */ | |
| 229 basic_block bb; | |
| 230 | |
| 111 | 231 /* We only create a template block for the first duplicated block in a |
| 232 jump threading path as we may need many duplicates of that block. | |
| 233 | |
| 234 The second duplicate block in a path is specific to that path. Creating | |
| 235 and sharing a template for that block is considerably more difficult. */ | |
| 0 | 236 basic_block template_block; |
| 237 | |
| 111 | 238 /* Blocks duplicated for the thread. */ |
| 239 bitmap duplicate_blocks; | |
| 240 | |
| 0 | 241 /* TRUE if we thread one or more jumps, FALSE otherwise. */ |
| 242 bool jumps_threaded; | |
| 111 | 243 |
| 244 /* When we have multiple paths through a joiner which reach different | |
| 245 final destinations, then we may need to correct for potential | |
| 246 profile insanities. */ | |
| 247 bool need_profile_correction; | |
| 0 | 248 }; |
| 249 | |
| 250 /* Passes which use the jump threading code register jump threading | |
| 251 opportunities as they are discovered. We keep the registered | |
| 252 jump threading opportunities in this vector as edge pairs | |
| 253 (original_edge, target_edge). */ | |
| 111 | 254 static vec<vec<jump_thread_edge *> *> paths; |
| 255 | |
| 256 /* When we start updating the CFG for threading, data necessary for jump | |
| 257 threading is attached to the AUX field for the incoming edge. Use these | |
| 258 macros to access the underlying structure attached to the AUX field. */ | |
| 259 #define THREAD_PATH(E) ((vec<jump_thread_edge *> *)(E)->aux) | |
| 0 | 260 |
| 261 /* Jump threading statistics. */ | |
| 262 | |
| 263 struct thread_stats_d | |
| 264 { | |
| 265 unsigned long num_threaded_edges; | |
| 266 }; | |
| 267 | |
| 268 struct thread_stats_d thread_stats; | |
| 269 | |
| 270 | |
| 271 /* Remove the last statement in block BB if it is a control statement | |
| 272 Also remove all outgoing edges except the edge which reaches DEST_BB. | |
| 273 If DEST_BB is NULL, then remove all outgoing edges. */ | |
| 274 | |
| 111 | 275 void |
| 0 | 276 remove_ctrl_stmt_and_useless_edges (basic_block bb, basic_block dest_bb) |
| 277 { | |
| 278 gimple_stmt_iterator gsi; | |
| 279 edge e; | |
| 280 edge_iterator ei; | |
| 281 | |
| 282 gsi = gsi_last_bb (bb); | |
| 283 | |
| 284 /* If the duplicate ends with a control statement, then remove it. | |
| 285 | |
| 286 Note that if we are duplicating the template block rather than the | |
| 287 original basic block, then the duplicate might not have any real | |
| 288 statements in it. */ | |
| 289 if (!gsi_end_p (gsi) | |
| 290 && gsi_stmt (gsi) | |
| 291 && (gimple_code (gsi_stmt (gsi)) == GIMPLE_COND | |
| 292 || gimple_code (gsi_stmt (gsi)) == GIMPLE_GOTO | |
| 293 || gimple_code (gsi_stmt (gsi)) == GIMPLE_SWITCH)) | |
| 294 gsi_remove (&gsi, true); | |
| 295 | |
| 296 for (ei = ei_start (bb->succs); (e = ei_safe_edge (ei)); ) | |
| 297 { | |
| 298 if (e->dest != dest_bb) | |
| 111 | 299 { |
| 300 free_dom_edge_info (e); | |
| 301 remove_edge (e); | |
| 302 } | |
| 0 | 303 else |
| 111 | 304 { |
| 305 e->probability = profile_probability::always (); | |
| 306 ei_next (&ei); | |
| 307 } | |
| 0 | 308 } |
| 111 | 309 |
| 310 /* If the remaining edge is a loop exit, there must have | |
| 311 a removed edge that was not a loop exit. | |
| 312 | |
| 313 In that case BB and possibly other blocks were previously | |
| 314 in the loop, but are now outside the loop. Thus, we need | |
| 315 to update the loop structures. */ | |
| 316 if (single_succ_p (bb) | |
| 317 && loop_outer (bb->loop_father) | |
| 318 && loop_exit_edge_p (bb->loop_father, single_succ_edge (bb))) | |
| 319 loops_state_set (LOOPS_NEED_FIXUP); | |
| 0 | 320 } |
| 321 | |
| 111 | 322 /* Create a duplicate of BB. Record the duplicate block in an array |
| 323 indexed by COUNT stored in RD. */ | |
| 0 | 324 |
| 325 static void | |
| 111 | 326 create_block_for_threading (basic_block bb, |
| 327 struct redirection_data *rd, | |
| 328 unsigned int count, | |
| 329 bitmap *duplicate_blocks) | |
| 0 | 330 { |
| 111 | 331 edge_iterator ei; |
| 332 edge e; | |
| 333 | |
| 0 | 334 /* We can use the generic block duplication code and simply remove |
| 335 the stuff we do not need. */ | |
| 111 | 336 rd->dup_blocks[count] = duplicate_block (bb, NULL, NULL); |
| 337 | |
| 338 FOR_EACH_EDGE (e, ei, rd->dup_blocks[count]->succs) | |
| 339 e->aux = NULL; | |
| 0 | 340 |
| 341 /* Zero out the profile, since the block is unreachable for now. */ | |
| 111 | 342 rd->dup_blocks[count]->frequency = 0; |
| 343 rd->dup_blocks[count]->count = profile_count::uninitialized (); | |
| 344 if (duplicate_blocks) | |
| 345 bitmap_set_bit (*duplicate_blocks, rd->dup_blocks[count]->index); | |
| 0 | 346 } |
| 347 | |
| 111 | 348 /* Main data structure to hold information for duplicates of BB. */ |
| 349 | |
| 350 static hash_table<redirection_data> *redirection_data; | |
| 0 | 351 |
| 352 /* Given an outgoing edge E lookup and return its entry in our hash table. | |
| 353 | |
| 354 If INSERT is true, then we insert the entry into the hash table if | |
| 355 it is not already present. INCOMING_EDGE is added to the list of incoming | |
| 356 edges associated with E in the hash table. */ | |
| 357 | |
| 358 static struct redirection_data * | |
| 111 | 359 lookup_redirection_data (edge e, enum insert_option insert) |
| 0 | 360 { |
| 111 | 361 struct redirection_data **slot; |
| 0 | 362 struct redirection_data *elt; |
| 111 | 363 vec<jump_thread_edge *> *path = THREAD_PATH (e); |
| 364 | |
| 365 /* Build a hash table element so we can see if E is already | |
| 0 | 366 in the table. */ |
| 367 elt = XNEW (struct redirection_data); | |
| 111 | 368 elt->path = path; |
| 369 elt->dup_blocks[0] = NULL; | |
| 370 elt->dup_blocks[1] = NULL; | |
| 0 | 371 elt->incoming_edges = NULL; |
| 372 | |
| 111 | 373 slot = redirection_data->find_slot (elt, insert); |
| 0 | 374 |
| 375 /* This will only happen if INSERT is false and the entry is not | |
| 376 in the hash table. */ | |
| 377 if (slot == NULL) | |
| 378 { | |
| 379 free (elt); | |
| 380 return NULL; | |
| 381 } | |
| 382 | |
| 383 /* This will only happen if E was not in the hash table and | |
| 384 INSERT is true. */ | |
| 385 if (*slot == NULL) | |
| 386 { | |
| 111 | 387 *slot = elt; |
| 0 | 388 elt->incoming_edges = XNEW (struct el); |
| 111 | 389 elt->incoming_edges->e = e; |
| 0 | 390 elt->incoming_edges->next = NULL; |
| 391 return elt; | |
| 392 } | |
| 393 /* E was in the hash table. */ | |
| 394 else | |
| 395 { | |
| 396 /* Free ELT as we do not need it anymore, we will extract the | |
| 397 relevant entry from the hash table itself. */ | |
| 398 free (elt); | |
| 399 | |
| 400 /* Get the entry stored in the hash table. */ | |
| 111 | 401 elt = *slot; |
| 0 | 402 |
| 403 /* If insertion was requested, then we need to add INCOMING_EDGE | |
| 404 to the list of incoming edges associated with E. */ | |
| 405 if (insert) | |
| 406 { | |
| 111 | 407 struct el *el = XNEW (struct el); |
| 0 | 408 el->next = elt->incoming_edges; |
| 111 | 409 el->e = e; |
| 0 | 410 elt->incoming_edges = el; |
| 411 } | |
| 412 | |
| 413 return elt; | |
| 414 } | |
| 415 } | |
| 416 | |
| 111 | 417 /* Similar to copy_phi_args, except that the PHI arg exists, it just |
| 418 does not have a value associated with it. */ | |
| 419 | |
| 420 static void | |
| 421 copy_phi_arg_into_existing_phi (edge src_e, edge tgt_e) | |
| 422 { | |
| 423 int src_idx = src_e->dest_idx; | |
| 424 int tgt_idx = tgt_e->dest_idx; | |
| 425 | |
| 426 /* Iterate over each PHI in e->dest. */ | |
| 427 for (gphi_iterator gsi = gsi_start_phis (src_e->dest), | |
| 428 gsi2 = gsi_start_phis (tgt_e->dest); | |
| 429 !gsi_end_p (gsi); | |
| 430 gsi_next (&gsi), gsi_next (&gsi2)) | |
| 431 { | |
| 432 gphi *src_phi = gsi.phi (); | |
| 433 gphi *dest_phi = gsi2.phi (); | |
| 434 tree val = gimple_phi_arg_def (src_phi, src_idx); | |
| 435 source_location locus = gimple_phi_arg_location (src_phi, src_idx); | |
| 436 | |
| 437 SET_PHI_ARG_DEF (dest_phi, tgt_idx, val); | |
| 438 gimple_phi_arg_set_location (dest_phi, tgt_idx, locus); | |
| 439 } | |
| 440 } | |
| 441 | |
| 442 /* Given ssa_name DEF, backtrack jump threading PATH from node IDX | |
| 443 to see if it has constant value in a flow sensitive manner. Set | |
| 444 LOCUS to location of the constant phi arg and return the value. | |
| 445 Return DEF directly if either PATH or idx is ZERO. */ | |
| 446 | |
| 447 static tree | |
| 448 get_value_locus_in_path (tree def, vec<jump_thread_edge *> *path, | |
| 449 basic_block bb, int idx, source_location *locus) | |
| 450 { | |
| 451 tree arg; | |
| 452 gphi *def_phi; | |
| 453 basic_block def_bb; | |
| 454 | |
| 455 if (path == NULL || idx == 0) | |
| 456 return def; | |
| 457 | |
| 458 def_phi = dyn_cast <gphi *> (SSA_NAME_DEF_STMT (def)); | |
| 459 if (!def_phi) | |
| 460 return def; | |
| 461 | |
| 462 def_bb = gimple_bb (def_phi); | |
| 463 /* Don't propagate loop invariants into deeper loops. */ | |
| 464 if (!def_bb || bb_loop_depth (def_bb) < bb_loop_depth (bb)) | |
| 465 return def; | |
| 466 | |
| 467 /* Backtrack jump threading path from IDX to see if def has constant | |
| 468 value. */ | |
| 469 for (int j = idx - 1; j >= 0; j--) | |
| 470 { | |
| 471 edge e = (*path)[j]->e; | |
| 472 if (e->dest == def_bb) | |
| 473 { | |
| 474 arg = gimple_phi_arg_def (def_phi, e->dest_idx); | |
| 475 if (is_gimple_min_invariant (arg)) | |
| 476 { | |
| 477 *locus = gimple_phi_arg_location (def_phi, e->dest_idx); | |
| 478 return arg; | |
| 479 } | |
| 480 break; | |
| 481 } | |
| 482 } | |
| 483 | |
| 484 return def; | |
| 485 } | |
| 486 | |
| 487 /* For each PHI in BB, copy the argument associated with SRC_E to TGT_E. | |
| 488 Try to backtrack jump threading PATH from node IDX to see if the arg | |
| 489 has constant value, copy constant value instead of argument itself | |
| 490 if yes. */ | |
| 491 | |
| 492 static void | |
| 493 copy_phi_args (basic_block bb, edge src_e, edge tgt_e, | |
| 494 vec<jump_thread_edge *> *path, int idx) | |
| 495 { | |
| 496 gphi_iterator gsi; | |
| 497 int src_indx = src_e->dest_idx; | |
| 498 | |
| 499 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 500 { | |
| 501 gphi *phi = gsi.phi (); | |
| 502 tree def = gimple_phi_arg_def (phi, src_indx); | |
| 503 source_location locus = gimple_phi_arg_location (phi, src_indx); | |
| 504 | |
| 505 if (TREE_CODE (def) == SSA_NAME | |
| 506 && !virtual_operand_p (gimple_phi_result (phi))) | |
| 507 def = get_value_locus_in_path (def, path, bb, idx, &locus); | |
| 508 | |
| 509 add_phi_arg (phi, def, tgt_e, locus); | |
| 510 } | |
| 511 } | |
| 512 | |
| 513 /* We have recently made a copy of ORIG_BB, including its outgoing | |
| 514 edges. The copy is NEW_BB. Every PHI node in every direct successor of | |
| 515 ORIG_BB has a new argument associated with edge from NEW_BB to the | |
| 516 successor. Initialize the PHI argument so that it is equal to the PHI | |
| 517 argument associated with the edge from ORIG_BB to the successor. | |
| 518 PATH and IDX are used to check if the new PHI argument has constant | |
| 519 value in a flow sensitive manner. */ | |
| 520 | |
| 521 static void | |
| 522 update_destination_phis (basic_block orig_bb, basic_block new_bb, | |
| 523 vec<jump_thread_edge *> *path, int idx) | |
| 524 { | |
| 525 edge_iterator ei; | |
| 526 edge e; | |
| 527 | |
| 528 FOR_EACH_EDGE (e, ei, orig_bb->succs) | |
| 529 { | |
| 530 edge e2 = find_edge (new_bb, e->dest); | |
| 531 copy_phi_args (e->dest, e, e2, path, idx); | |
| 532 } | |
| 533 } | |
| 534 | |
| 0 | 535 /* Given a duplicate block and its single destination (both stored |
| 536 in RD). Create an edge between the duplicate and its single | |
| 537 destination. | |
| 538 | |
| 539 Add an additional argument to any PHI nodes at the single | |
| 111 | 540 destination. IDX is the start node in jump threading path |
| 541 we start to check to see if the new PHI argument has constant | |
| 542 value along the jump threading path. */ | |
| 0 | 543 |
| 544 static void | |
| 111 | 545 create_edge_and_update_destination_phis (struct redirection_data *rd, |
| 546 basic_block bb, int idx) | |
| 0 | 547 { |
| 111 | 548 edge e = make_single_succ_edge (bb, rd->path->last ()->e->dest, EDGE_FALLTHRU); |
| 0 | 549 |
| 550 rescan_loop_exit (e, true, false); | |
| 111 | 551 |
| 552 /* We used to copy the thread path here. That was added in 2007 | |
| 553 and dutifully updated through the representation changes in 2013. | |
| 554 | |
| 555 In 2013 we added code to thread from an interior node through | |
| 556 the backedge to another interior node. That runs after the code | |
| 557 to thread through loop headers from outside the loop. | |
| 558 | |
| 559 The latter may delete edges in the CFG, including those | |
| 560 which appeared in the jump threading path we copied here. Thus | |
| 561 we'd end up using a dangling pointer. | |
| 562 | |
| 563 After reviewing the 2007/2011 code, I can't see how anything | |
| 564 depended on copying the AUX field and clearly copying the jump | |
| 565 threading path is problematical due to embedded edge pointers. | |
| 566 It has been removed. */ | |
| 567 e->aux = NULL; | |
| 0 | 568 |
| 569 /* If there are any PHI nodes at the destination of the outgoing edge | |
| 570 from the duplicate block, then we will need to add a new argument | |
| 571 to them. The argument should have the same value as the argument | |
| 572 associated with the outgoing edge stored in RD. */ | |
| 111 | 573 copy_phi_args (e->dest, rd->path->last ()->e, e, rd->path, idx); |
| 574 } | |
| 575 | |
| 576 /* Look through PATH beginning at START and return TRUE if there are | |
| 577 any additional blocks that need to be duplicated. Otherwise, | |
| 578 return FALSE. */ | |
| 579 static bool | |
| 580 any_remaining_duplicated_blocks (vec<jump_thread_edge *> *path, | |
| 581 unsigned int start) | |
| 582 { | |
| 583 for (unsigned int i = start + 1; i < path->length (); i++) | |
| 584 { | |
| 585 if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK | |
| 586 || (*path)[i]->type == EDGE_COPY_SRC_BLOCK) | |
| 587 return true; | |
| 588 } | |
| 589 return false; | |
| 590 } | |
| 591 | |
| 592 | |
| 593 /* Compute the amount of profile count/frequency coming into the jump threading | |
| 594 path stored in RD that we are duplicating, returned in PATH_IN_COUNT_PTR and | |
| 595 PATH_IN_FREQ_PTR, as well as the amount of counts flowing out of the | |
| 596 duplicated path, returned in PATH_OUT_COUNT_PTR. LOCAL_INFO is used to | |
| 597 identify blocks duplicated for jump threading, which have duplicated | |
| 598 edges that need to be ignored in the analysis. Return true if path contains | |
| 599 a joiner, false otherwise. | |
| 600 | |
| 601 In the non-joiner case, this is straightforward - all the counts/frequency | |
| 602 flowing into the jump threading path should flow through the duplicated | |
| 603 block and out of the duplicated path. | |
| 604 | |
| 605 In the joiner case, it is very tricky. Some of the counts flowing into | |
| 606 the original path go offpath at the joiner. The problem is that while | |
| 607 we know how much total count goes off-path in the original control flow, | |
| 608 we don't know how many of the counts corresponding to just the jump | |
| 609 threading path go offpath at the joiner. | |
| 610 | |
| 611 For example, assume we have the following control flow and identified | |
| 612 jump threading paths: | |
| 613 | |
| 614 A B C | |
| 615 \ | / | |
| 616 Ea \ |Eb / Ec | |
| 617 \ | / | |
| 618 v v v | |
| 619 J <-- Joiner | |
| 620 / \ | |
| 621 Eoff/ \Eon | |
| 622 / \ | |
| 623 v v | |
| 624 Soff Son <--- Normal | |
| 625 /\ | |
| 626 Ed/ \ Ee | |
| 627 / \ | |
| 628 v v | |
| 629 D E | |
| 630 | |
| 631 Jump threading paths: A -> J -> Son -> D (path 1) | |
| 632 C -> J -> Son -> E (path 2) | |
| 633 | |
| 634 Note that the control flow could be more complicated: | |
| 635 - Each jump threading path may have more than one incoming edge. I.e. A and | |
| 636 Ea could represent multiple incoming blocks/edges that are included in | |
| 637 path 1. | |
| 638 - There could be EDGE_NO_COPY_SRC_BLOCK edges after the joiner (either | |
| 639 before or after the "normal" copy block). These are not duplicated onto | |
| 640 the jump threading path, as they are single-successor. | |
| 641 - Any of the blocks along the path may have other incoming edges that | |
| 642 are not part of any jump threading path, but add profile counts along | |
| 643 the path. | |
| 644 | |
| 645 In the above example, after all jump threading is complete, we will | |
| 646 end up with the following control flow: | |
| 647 | |
| 648 A B C | |
| 649 | | | | |
| 650 Ea| |Eb |Ec | |
| 651 | | | | |
| 652 v v v | |
| 653 Ja J Jc | |
| 654 / \ / \Eon' / \ | |
| 655 Eona/ \ ---/---\-------- \Eonc | |
| 656 / \ / / \ \ | |
| 657 v v v v v | |
| 658 Sona Soff Son Sonc | |
| 659 \ /\ / | |
| 660 \___________ / \ _____/ | |
| 661 \ / \/ | |
| 662 vv v | |
| 663 D E | |
| 664 | |
| 665 The main issue to notice here is that when we are processing path 1 | |
| 666 (A->J->Son->D) we need to figure out the outgoing edge weights to | |
| 667 the duplicated edges Ja->Sona and Ja->Soff, while ensuring that the | |
| 668 sum of the incoming weights to D remain Ed. The problem with simply | |
| 669 assuming that Ja (and Jc when processing path 2) has the same outgoing | |
| 670 probabilities to its successors as the original block J, is that after | |
| 671 all paths are processed and other edges/counts removed (e.g. none | |
| 672 of Ec will reach D after processing path 2), we may end up with not | |
| 673 enough count flowing along duplicated edge Sona->D. | |
| 674 | |
| 675 Therefore, in the case of a joiner, we keep track of all counts | |
| 676 coming in along the current path, as well as from predecessors not | |
| 677 on any jump threading path (Eb in the above example). While we | |
| 678 first assume that the duplicated Eona for Ja->Sona has the same | |
| 679 probability as the original, we later compensate for other jump | |
| 680 threading paths that may eliminate edges. We do that by keep track | |
| 681 of all counts coming into the original path that are not in a jump | |
| 682 thread (Eb in the above example, but as noted earlier, there could | |
| 683 be other predecessors incoming to the path at various points, such | |
| 684 as at Son). Call this cumulative non-path count coming into the path | |
| 685 before D as Enonpath. We then ensure that the count from Sona->D is as at | |
| 686 least as big as (Ed - Enonpath), but no bigger than the minimum | |
| 687 weight along the jump threading path. The probabilities of both the | |
| 688 original and duplicated joiner block J and Ja will be adjusted | |
| 689 accordingly after the updates. */ | |
| 690 | |
| 691 static bool | |
| 692 compute_path_counts (struct redirection_data *rd, | |
| 693 ssa_local_info_t *local_info, | |
| 694 profile_count *path_in_count_ptr, | |
| 695 profile_count *path_out_count_ptr, | |
| 696 int *path_in_freq_ptr) | |
| 697 { | |
| 698 edge e = rd->incoming_edges->e; | |
| 699 vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
| 700 edge elast = path->last ()->e; | |
| 701 profile_count nonpath_count = profile_count::zero (); | |
| 702 bool has_joiner = false; | |
| 703 profile_count path_in_count = profile_count::zero (); | |
| 704 int path_in_freq = 0; | |
| 705 | |
| 706 /* Start by accumulating incoming edge counts to the path's first bb | |
| 707 into a couple buckets: | |
| 708 path_in_count: total count of incoming edges that flow into the | |
| 709 current path. | |
| 710 nonpath_count: total count of incoming edges that are not | |
| 711 flowing along *any* path. These are the counts | |
| 712 that will still flow along the original path after | |
| 713 all path duplication is done by potentially multiple | |
| 714 calls to this routine. | |
| 715 (any other incoming edge counts are for a different jump threading | |
| 716 path that will be handled by a later call to this routine.) | |
| 717 To make this easier, start by recording all incoming edges that flow into | |
| 718 the current path in a bitmap. We could add up the path's incoming edge | |
| 719 counts here, but we still need to walk all the first bb's incoming edges | |
| 720 below to add up the counts of the other edges not included in this jump | |
| 721 threading path. */ | |
| 722 struct el *next, *el; | |
| 723 auto_bitmap in_edge_srcs; | |
| 724 for (el = rd->incoming_edges; el; el = next) | |
| 725 { | |
| 726 next = el->next; | |
| 727 bitmap_set_bit (in_edge_srcs, el->e->src->index); | |
| 728 } | |
| 729 edge ein; | |
| 730 edge_iterator ei; | |
| 731 FOR_EACH_EDGE (ein, ei, e->dest->preds) | |
| 732 { | |
| 733 vec<jump_thread_edge *> *ein_path = THREAD_PATH (ein); | |
| 734 /* Simply check the incoming edge src against the set captured above. */ | |
| 735 if (ein_path | |
| 736 && bitmap_bit_p (in_edge_srcs, (*ein_path)[0]->e->src->index)) | |
| 737 { | |
| 738 /* It is necessary but not sufficient that the last path edges | |
| 739 are identical. There may be different paths that share the | |
| 740 same last path edge in the case where the last edge has a nocopy | |
| 741 source block. */ | |
| 742 gcc_assert (ein_path->last ()->e == elast); | |
| 743 path_in_count += ein->count (); | |
| 744 path_in_freq += EDGE_FREQUENCY (ein); | |
| 745 } | |
| 746 else if (!ein_path) | |
| 747 { | |
| 748 /* Keep track of the incoming edges that are not on any jump-threading | |
| 749 path. These counts will still flow out of original path after all | |
| 750 jump threading is complete. */ | |
| 751 nonpath_count += ein->count (); | |
| 752 } | |
| 753 } | |
| 754 | |
| 755 /* This is needed due to insane incoming frequencies. */ | |
| 756 if (path_in_freq > BB_FREQ_MAX) | |
| 757 path_in_freq = BB_FREQ_MAX; | |
| 758 | |
| 759 /* Now compute the fraction of the total count coming into the first | |
| 760 path bb that is from the current threading path. */ | |
| 761 profile_count total_count = e->dest->count; | |
| 762 /* Handle incoming profile insanities. */ | |
| 763 if (total_count < path_in_count) | |
| 764 path_in_count = total_count; | |
| 765 profile_probability onpath_scale = path_in_count.probability_in (total_count); | |
| 766 | |
| 767 /* Walk the entire path to do some more computation in order to estimate | |
| 768 how much of the path_in_count will flow out of the duplicated threading | |
| 769 path. In the non-joiner case this is straightforward (it should be | |
| 770 the same as path_in_count, although we will handle incoming profile | |
| 771 insanities by setting it equal to the minimum count along the path). | |
| 772 | |
| 773 In the joiner case, we need to estimate how much of the path_in_count | |
| 774 will stay on the threading path after the joiner's conditional branch. | |
| 775 We don't really know for sure how much of the counts | |
| 776 associated with this path go to each successor of the joiner, but we'll | |
| 777 estimate based on the fraction of the total count coming into the path | |
| 778 bb was from the threading paths (computed above in onpath_scale). | |
| 779 Afterwards, we will need to do some fixup to account for other threading | |
| 780 paths and possible profile insanities. | |
| 781 | |
| 782 In order to estimate the joiner case's counts we also need to update | |
| 783 nonpath_count with any additional counts coming into the path. Other | |
| 784 blocks along the path may have additional predecessors from outside | |
| 785 the path. */ | |
| 786 profile_count path_out_count = path_in_count; | |
| 787 profile_count min_path_count = path_in_count; | |
| 788 for (unsigned int i = 1; i < path->length (); i++) | |
| 789 { | |
| 790 edge epath = (*path)[i]->e; | |
| 791 profile_count cur_count = epath->count (); | |
| 792 if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
| 793 { | |
| 794 has_joiner = true; | |
| 795 cur_count = cur_count.apply_probability (onpath_scale); | |
| 796 } | |
| 797 /* In the joiner case we need to update nonpath_count for any edges | |
| 798 coming into the path that will contribute to the count flowing | |
| 799 into the path successor. */ | |
| 800 if (has_joiner && epath != elast) | |
| 801 { | |
| 802 /* Look for other incoming edges after joiner. */ | |
| 803 FOR_EACH_EDGE (ein, ei, epath->dest->preds) | |
| 804 { | |
| 805 if (ein != epath | |
| 806 /* Ignore in edges from blocks we have duplicated for a | |
| 807 threading path, which have duplicated edge counts until | |
| 808 they are redirected by an invocation of this routine. */ | |
| 809 && !bitmap_bit_p (local_info->duplicate_blocks, | |
| 810 ein->src->index)) | |
| 811 nonpath_count += ein->count (); | |
| 812 } | |
| 813 } | |
| 814 if (cur_count < path_out_count) | |
| 815 path_out_count = cur_count; | |
| 816 if (epath->count () < min_path_count) | |
| 817 min_path_count = epath->count (); | |
| 818 } | |
| 819 | |
| 820 /* We computed path_out_count above assuming that this path targeted | |
| 821 the joiner's on-path successor with the same likelihood as it | |
| 822 reached the joiner. However, other thread paths through the joiner | |
| 823 may take a different path through the normal copy source block | |
| 824 (i.e. they have a different elast), meaning that they do not | |
| 825 contribute any counts to this path's elast. As a result, it may | |
| 826 turn out that this path must have more count flowing to the on-path | |
| 827 successor of the joiner. Essentially, all of this path's elast | |
| 828 count must be contributed by this path and any nonpath counts | |
| 829 (since any path through the joiner with a different elast will not | |
| 830 include a copy of this elast in its duplicated path). | |
| 831 So ensure that this path's path_out_count is at least the | |
| 832 difference between elast->count () and nonpath_count. Otherwise the edge | |
| 833 counts after threading will not be sane. */ | |
| 834 if (local_info->need_profile_correction | |
| 835 && has_joiner && path_out_count < elast->count () - nonpath_count) | |
| 836 { | |
| 837 path_out_count = elast->count () - nonpath_count; | |
| 838 /* But neither can we go above the minimum count along the path | |
| 839 we are duplicating. This can be an issue due to profile | |
| 840 insanities coming in to this pass. */ | |
| 841 if (path_out_count > min_path_count) | |
| 842 path_out_count = min_path_count; | |
| 843 } | |
| 844 | |
| 845 *path_in_count_ptr = path_in_count; | |
| 846 *path_out_count_ptr = path_out_count; | |
| 847 *path_in_freq_ptr = path_in_freq; | |
| 848 return has_joiner; | |
| 849 } | |
| 850 | |
| 851 | |
| 852 /* Update the counts and frequencies for both an original path | |
| 853 edge EPATH and its duplicate EDUP. The duplicate source block | |
| 854 will get a count/frequency of PATH_IN_COUNT and PATH_IN_FREQ, | |
| 855 and the duplicate edge EDUP will have a count of PATH_OUT_COUNT. */ | |
| 856 static void | |
| 857 update_profile (edge epath, edge edup, profile_count path_in_count, | |
| 858 profile_count path_out_count, int path_in_freq) | |
| 859 { | |
| 860 if (!(path_in_count > 0)) | |
| 861 return; | |
| 862 | |
| 863 /* First update the duplicated block's count / frequency. */ | |
| 864 if (edup) | |
| 0 | 865 { |
| 111 | 866 basic_block dup_block = edup->src; |
| 867 | |
| 868 /* Edup's count is reduced by path_out_count. We need to redistribute | |
| 869 probabilities to the remaining edges. */ | |
| 870 | |
| 871 edge esucc; | |
| 872 edge_iterator ei; | |
| 873 profile_probability edup_prob | |
| 874 = path_out_count.probability_in (path_in_count); | |
| 875 | |
| 876 /* Either scale up or down the remaining edges. | |
| 877 probabilities are always in range <0,1> and thus we can't do | |
| 878 both by same loop. */ | |
| 879 if (edup->probability > edup_prob) | |
| 880 { | |
| 881 profile_probability rev_scale | |
| 882 = (profile_probability::always () - edup->probability) | |
| 883 / (profile_probability::always () - edup_prob); | |
| 884 FOR_EACH_EDGE (esucc, ei, dup_block->succs) | |
| 885 if (esucc != edup) | |
| 886 esucc->probability /= rev_scale; | |
| 887 } | |
| 888 else if (edup->probability < edup_prob) | |
| 889 { | |
| 890 profile_probability scale | |
| 891 = (profile_probability::always () - edup_prob) | |
| 892 / (profile_probability::always () - edup->probability); | |
| 893 FOR_EACH_EDGE (esucc, ei, dup_block->succs) | |
| 894 if (esucc != edup) | |
| 895 esucc->probability *= scale; | |
| 896 } | |
| 897 edup->probability = edup_prob; | |
| 898 | |
| 899 /* FIXME once freqs_to_counts is dropped re-enable this check. */ | |
| 900 gcc_assert (!dup_block->count.initialized_p () || 1); | |
| 901 gcc_assert (dup_block->frequency == 0); | |
| 902 dup_block->count = path_in_count; | |
| 903 dup_block->frequency = path_in_freq; | |
| 904 } | |
| 905 | |
| 906 profile_count final_count = epath->count () - path_out_count; | |
| 907 | |
| 908 /* Now update the original block's count and frequency in the | |
| 909 opposite manner - remove the counts/freq that will flow | |
| 910 into the duplicated block. Handle underflow due to precision/ | |
| 911 rounding issues. */ | |
| 912 epath->src->count -= path_in_count; | |
| 913 epath->src->frequency -= path_in_freq; | |
| 914 if (epath->src->frequency < 0) | |
| 915 epath->src->frequency = 0; | |
| 916 | |
| 917 /* Next update this path edge's original and duplicated counts. We know | |
| 918 that the duplicated path will have path_out_count flowing | |
| 919 out of it (in the joiner case this is the count along the duplicated path | |
| 920 out of the duplicated joiner). This count can then be removed from the | |
| 921 original path edge. */ | |
| 922 if (epath->src->count > 0) | |
| 923 { | |
| 924 edge esucc; | |
| 925 edge_iterator ei; | |
| 926 profile_probability epath_prob = final_count.probability_in (epath->src->count); | |
| 927 | |
| 928 if (epath->probability > epath_prob) | |
| 929 { | |
| 930 profile_probability rev_scale | |
| 931 = (profile_probability::always () - epath->probability) | |
| 932 / (profile_probability::always () - epath_prob); | |
| 933 FOR_EACH_EDGE (esucc, ei, epath->src->succs) | |
| 934 if (esucc != epath) | |
| 935 esucc->probability /= rev_scale; | |
| 936 } | |
| 937 else if (epath->probability < epath_prob) | |
| 938 { | |
| 939 profile_probability scale | |
| 940 = (profile_probability::always () - epath_prob) | |
| 941 / (profile_probability::always () - epath->probability); | |
| 942 FOR_EACH_EDGE (esucc, ei, epath->src->succs) | |
| 943 if (esucc != epath) | |
| 944 esucc->probability *= scale; | |
| 945 } | |
| 946 epath->probability = epath_prob; | |
| 947 } | |
| 948 } | |
| 949 | |
| 950 | |
| 951 /* Check if the paths through RD all have estimated frequencies but zero | |
| 952 profile counts. This is more accurate than checking the entry block | |
| 953 for a zero profile count, since profile insanities sometimes creep in. */ | |
| 954 | |
| 955 static bool | |
| 956 estimated_freqs_path (struct redirection_data *rd) | |
| 957 { | |
| 958 edge e = rd->incoming_edges->e; | |
| 959 vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
| 960 edge ein; | |
| 961 edge_iterator ei; | |
| 962 bool non_zero_freq = false; | |
| 963 FOR_EACH_EDGE (ein, ei, e->dest->preds) | |
| 964 { | |
| 965 if (ein->count () > 0) | |
| 966 return false; | |
| 967 non_zero_freq |= ein->src->frequency != 0; | |
| 968 } | |
| 969 | |
| 970 for (unsigned int i = 1; i < path->length (); i++) | |
| 971 { | |
| 972 edge epath = (*path)[i]->e; | |
| 973 if (epath->src->count > 0) | |
| 974 return false; | |
| 975 non_zero_freq |= epath->src->frequency != 0; | |
| 976 edge esucc; | |
| 977 FOR_EACH_EDGE (esucc, ei, epath->src->succs) | |
| 978 { | |
| 979 if (esucc->count () > 0) | |
| 980 return false; | |
| 981 non_zero_freq |= esucc->src->frequency != 0; | |
| 982 } | |
| 983 } | |
| 984 return non_zero_freq; | |
| 985 } | |
| 986 | |
| 987 | |
| 988 /* Invoked for routines that have guessed frequencies and no profile | |
| 989 counts to record the block and edge frequencies for paths through RD | |
| 990 in the profile count fields of those blocks and edges. This is because | |
| 991 ssa_fix_duplicate_block_edges incrementally updates the block and | |
| 992 edge counts as edges are redirected, and it is difficult to do that | |
| 993 for edge frequencies which are computed on the fly from the source | |
| 994 block frequency and probability. When a block frequency is updated | |
| 995 its outgoing edge frequencies are affected and become difficult to | |
| 996 adjust. */ | |
| 997 | |
| 998 static void | |
| 999 freqs_to_counts_path (struct redirection_data *rd) | |
| 1000 { | |
| 1001 edge e = rd->incoming_edges->e; | |
| 1002 vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
| 1003 edge ein; | |
| 1004 edge_iterator ei; | |
| 1005 | |
| 1006 FOR_EACH_EDGE (ein, ei, e->dest->preds) | |
| 1007 ein->src->count = profile_count::from_gcov_type | |
| 1008 (ein->src->frequency * REG_BR_PROB_BASE); | |
| 1009 for (unsigned int i = 1; i < path->length (); i++) | |
| 1010 { | |
| 1011 edge epath = (*path)[i]->e; | |
| 1012 /* Scale up the frequency by REG_BR_PROB_BASE, to avoid rounding | |
| 1013 errors applying the edge probability when the frequencies are very | |
| 1014 small. */ | |
| 1015 epath->src->count = | |
| 1016 profile_count::from_gcov_type | |
| 1017 (epath->src->frequency * REG_BR_PROB_BASE); | |
| 0 | 1018 } |
| 1019 } | |
| 1020 | |
| 111 | 1021 |
| 1022 /* For routines that have guessed frequencies and no profile counts, where we | |
| 1023 used freqs_to_counts_path to record block and edge frequencies for paths | |
| 1024 through RD, we clear the counts after completing all updates for RD. | |
| 1025 The updates in ssa_fix_duplicate_block_edges are based off the count fields, | |
| 1026 but the block frequencies and edge probabilities were updated as well, | |
| 1027 so we can simply clear the count fields. */ | |
| 1028 | |
| 1029 static void | |
| 1030 clear_counts_path (struct redirection_data *rd) | |
| 1031 { | |
| 1032 edge e = rd->incoming_edges->e; | |
| 1033 vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
| 1034 profile_count val = profile_count::uninitialized (); | |
| 1035 if (profile_status_for_fn (cfun) == PROFILE_READ) | |
| 1036 val = profile_count::zero (); | |
| 1037 | |
| 1038 edge ein; | |
| 1039 edge_iterator ei; | |
| 1040 | |
| 1041 FOR_EACH_EDGE (ein, ei, e->dest->preds) | |
| 1042 ein->src->count = val; | |
| 1043 | |
| 1044 /* First clear counts along original path. */ | |
| 1045 for (unsigned int i = 1; i < path->length (); i++) | |
| 1046 { | |
| 1047 edge epath = (*path)[i]->e; | |
| 1048 epath->src->count = val; | |
| 1049 } | |
| 1050 /* Also need to clear the counts along duplicated path. */ | |
| 1051 for (unsigned int i = 0; i < 2; i++) | |
| 1052 { | |
| 1053 basic_block dup = rd->dup_blocks[i]; | |
| 1054 if (!dup) | |
| 1055 continue; | |
| 1056 dup->count = val; | |
| 1057 } | |
| 1058 } | |
| 1059 | |
| 1060 /* Wire up the outgoing edges from the duplicate blocks and | |
| 1061 update any PHIs as needed. Also update the profile counts | |
| 1062 on the original and duplicate blocks and edges. */ | |
| 1063 void | |
| 1064 ssa_fix_duplicate_block_edges (struct redirection_data *rd, | |
| 1065 ssa_local_info_t *local_info) | |
| 1066 { | |
| 1067 bool multi_incomings = (rd->incoming_edges->next != NULL); | |
| 1068 edge e = rd->incoming_edges->e; | |
| 1069 vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
| 1070 edge elast = path->last ()->e; | |
| 1071 profile_count path_in_count = profile_count::zero (); | |
| 1072 profile_count path_out_count = profile_count::zero (); | |
| 1073 int path_in_freq = 0; | |
| 1074 | |
| 1075 /* This routine updates profile counts, frequencies, and probabilities | |
| 1076 incrementally. Since it is difficult to do the incremental updates | |
| 1077 using frequencies/probabilities alone, for routines without profile | |
| 1078 data we first take a snapshot of the existing block and edge frequencies | |
| 1079 by copying them into the empty profile count fields. These counts are | |
| 1080 then used to do the incremental updates, and cleared at the end of this | |
| 1081 routine. If the function is marked as having a profile, we still check | |
| 1082 to see if the paths through RD are using estimated frequencies because | |
| 1083 the routine had zero profile counts. */ | |
| 1084 bool do_freqs_to_counts = (profile_status_for_fn (cfun) != PROFILE_READ | |
| 1085 || estimated_freqs_path (rd)); | |
| 1086 if (do_freqs_to_counts) | |
| 1087 freqs_to_counts_path (rd); | |
| 1088 | |
| 1089 /* First determine how much profile count to move from original | |
| 1090 path to the duplicate path. This is tricky in the presence of | |
| 1091 a joiner (see comments for compute_path_counts), where some portion | |
| 1092 of the path's counts will flow off-path from the joiner. In the | |
| 1093 non-joiner case the path_in_count and path_out_count should be the | |
| 1094 same. */ | |
| 1095 bool has_joiner = compute_path_counts (rd, local_info, | |
| 1096 &path_in_count, &path_out_count, | |
| 1097 &path_in_freq); | |
| 1098 | |
| 1099 int cur_path_freq = path_in_freq; | |
| 1100 for (unsigned int count = 0, i = 1; i < path->length (); i++) | |
| 1101 { | |
| 1102 edge epath = (*path)[i]->e; | |
| 1103 | |
| 1104 /* If we were threading through an joiner block, then we want | |
| 1105 to keep its control statement and redirect an outgoing edge. | |
| 1106 Else we want to remove the control statement & edges, then create | |
| 1107 a new outgoing edge. In both cases we may need to update PHIs. */ | |
| 1108 if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
| 1109 { | |
| 1110 edge victim; | |
| 1111 edge e2; | |
| 1112 | |
| 1113 gcc_assert (has_joiner); | |
| 1114 | |
| 1115 /* This updates the PHIs at the destination of the duplicate | |
| 1116 block. Pass 0 instead of i if we are threading a path which | |
| 1117 has multiple incoming edges. */ | |
| 1118 update_destination_phis (local_info->bb, rd->dup_blocks[count], | |
| 1119 path, multi_incomings ? 0 : i); | |
| 1120 | |
| 1121 /* Find the edge from the duplicate block to the block we're | |
| 1122 threading through. That's the edge we want to redirect. */ | |
| 1123 victim = find_edge (rd->dup_blocks[count], (*path)[i]->e->dest); | |
| 1124 | |
| 1125 /* If there are no remaining blocks on the path to duplicate, | |
| 1126 then redirect VICTIM to the final destination of the jump | |
| 1127 threading path. */ | |
| 1128 if (!any_remaining_duplicated_blocks (path, i)) | |
| 1129 { | |
| 1130 e2 = redirect_edge_and_branch (victim, elast->dest); | |
| 1131 /* If we redirected the edge, then we need to copy PHI arguments | |
| 1132 at the target. If the edge already existed (e2 != victim | |
| 1133 case), then the PHIs in the target already have the correct | |
| 1134 arguments. */ | |
| 1135 if (e2 == victim) | |
| 1136 copy_phi_args (e2->dest, elast, e2, | |
| 1137 path, multi_incomings ? 0 : i); | |
| 1138 } | |
| 1139 else | |
| 1140 { | |
| 1141 /* Redirect VICTIM to the next duplicated block in the path. */ | |
| 1142 e2 = redirect_edge_and_branch (victim, rd->dup_blocks[count + 1]); | |
| 1143 | |
| 1144 /* We need to update the PHIs in the next duplicated block. We | |
| 1145 want the new PHI args to have the same value as they had | |
| 1146 in the source of the next duplicate block. | |
| 1147 | |
| 1148 Thus, we need to know which edge we traversed into the | |
| 1149 source of the duplicate. Furthermore, we may have | |
| 1150 traversed many edges to reach the source of the duplicate. | |
| 1151 | |
| 1152 Walk through the path starting at element I until we | |
| 1153 hit an edge marked with EDGE_COPY_SRC_BLOCK. We want | |
| 1154 the edge from the prior element. */ | |
| 1155 for (unsigned int j = i + 1; j < path->length (); j++) | |
| 1156 { | |
| 1157 if ((*path)[j]->type == EDGE_COPY_SRC_BLOCK) | |
| 1158 { | |
| 1159 copy_phi_arg_into_existing_phi ((*path)[j - 1]->e, e2); | |
| 1160 break; | |
| 1161 } | |
| 1162 } | |
| 1163 } | |
| 1164 | |
| 1165 /* Update the counts and frequency of both the original block | |
| 1166 and path edge, and the duplicates. The path duplicate's | |
| 1167 incoming count and frequency are the totals for all edges | |
| 1168 incoming to this jump threading path computed earlier. | |
| 1169 And we know that the duplicated path will have path_out_count | |
| 1170 flowing out of it (i.e. along the duplicated path out of the | |
| 1171 duplicated joiner). */ | |
| 1172 update_profile (epath, e2, path_in_count, path_out_count, | |
| 1173 path_in_freq); | |
| 1174 | |
| 1175 /* Record the frequency flowing to the downstream duplicated | |
| 1176 path blocks. */ | |
| 1177 cur_path_freq = EDGE_FREQUENCY (e2); | |
| 1178 } | |
| 1179 else if ((*path)[i]->type == EDGE_COPY_SRC_BLOCK) | |
| 1180 { | |
| 1181 remove_ctrl_stmt_and_useless_edges (rd->dup_blocks[count], NULL); | |
| 1182 create_edge_and_update_destination_phis (rd, rd->dup_blocks[count], | |
| 1183 multi_incomings ? 0 : i); | |
| 1184 if (count == 1) | |
| 1185 single_succ_edge (rd->dup_blocks[1])->aux = NULL; | |
| 1186 | |
| 1187 /* Update the counts and frequency of both the original block | |
| 1188 and path edge, and the duplicates. Since we are now after | |
| 1189 any joiner that may have existed on the path, the count | |
| 1190 flowing along the duplicated threaded path is path_out_count. | |
| 1191 If we didn't have a joiner, then cur_path_freq was the sum | |
| 1192 of the total frequencies along all incoming edges to the | |
| 1193 thread path (path_in_freq). If we had a joiner, it would have | |
| 1194 been updated at the end of that handling to the edge frequency | |
| 1195 along the duplicated joiner path edge. */ | |
| 1196 update_profile (epath, EDGE_SUCC (rd->dup_blocks[count], 0), | |
| 1197 path_out_count, path_out_count, cur_path_freq); | |
| 1198 } | |
| 1199 else | |
| 1200 { | |
| 1201 /* No copy case. In this case we don't have an equivalent block | |
| 1202 on the duplicated thread path to update, but we do need | |
| 1203 to remove the portion of the counts/freqs that were moved | |
| 1204 to the duplicated path from the counts/freqs flowing through | |
| 1205 this block on the original path. Since all the no-copy edges | |
| 1206 are after any joiner, the removed count is the same as | |
| 1207 path_out_count. | |
| 1208 | |
| 1209 If we didn't have a joiner, then cur_path_freq was the sum | |
| 1210 of the total frequencies along all incoming edges to the | |
| 1211 thread path (path_in_freq). If we had a joiner, it would have | |
| 1212 been updated at the end of that handling to the edge frequency | |
| 1213 along the duplicated joiner path edge. */ | |
| 1214 update_profile (epath, NULL, path_out_count, path_out_count, | |
| 1215 cur_path_freq); | |
| 1216 } | |
| 1217 | |
| 1218 /* Increment the index into the duplicated path when we processed | |
| 1219 a duplicated block. */ | |
| 1220 if ((*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK | |
| 1221 || (*path)[i]->type == EDGE_COPY_SRC_BLOCK) | |
| 1222 { | |
| 1223 count++; | |
| 1224 } | |
| 1225 } | |
| 1226 | |
| 1227 /* Done with all profile and frequency updates, clear counts if they | |
| 1228 were copied. */ | |
| 1229 if (do_freqs_to_counts) | |
| 1230 clear_counts_path (rd); | |
| 1231 } | |
| 1232 | |
| 0 | 1233 /* Hash table traversal callback routine to create duplicate blocks. */ |
| 1234 | |
| 111 | 1235 int |
| 1236 ssa_create_duplicates (struct redirection_data **slot, | |
| 1237 ssa_local_info_t *local_info) | |
| 0 | 1238 { |
| 111 | 1239 struct redirection_data *rd = *slot; |
| 1240 | |
| 1241 /* The second duplicated block in a jump threading path is specific | |
| 1242 to the path. So it gets stored in RD rather than in LOCAL_DATA. | |
| 1243 | |
| 1244 Each time we're called, we have to look through the path and see | |
| 1245 if a second block needs to be duplicated. | |
| 1246 | |
| 1247 Note the search starts with the third edge on the path. The first | |
| 1248 edge is the incoming edge, the second edge always has its source | |
| 1249 duplicated. Thus we start our search with the third edge. */ | |
| 1250 vec<jump_thread_edge *> *path = rd->path; | |
| 1251 for (unsigned int i = 2; i < path->length (); i++) | |
| 1252 { | |
| 1253 if ((*path)[i]->type == EDGE_COPY_SRC_BLOCK | |
| 1254 || (*path)[i]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
| 1255 { | |
| 1256 create_block_for_threading ((*path)[i]->e->src, rd, 1, | |
| 1257 &local_info->duplicate_blocks); | |
| 1258 break; | |
| 1259 } | |
| 1260 } | |
| 0 | 1261 |
| 1262 /* Create a template block if we have not done so already. Otherwise | |
| 1263 use the template to create a new block. */ | |
| 1264 if (local_info->template_block == NULL) | |
| 1265 { | |
| 111 | 1266 create_block_for_threading ((*path)[1]->e->src, rd, 0, |
| 1267 &local_info->duplicate_blocks); | |
| 1268 local_info->template_block = rd->dup_blocks[0]; | |
| 0 | 1269 |
| 1270 /* We do not create any outgoing edges for the template. We will | |
| 1271 take care of that in a later traversal. That way we do not | |
| 1272 create edges that are going to just be deleted. */ | |
| 1273 } | |
| 1274 else | |
| 1275 { | |
| 111 | 1276 create_block_for_threading (local_info->template_block, rd, 0, |
| 1277 &local_info->duplicate_blocks); | |
| 0 | 1278 |
| 1279 /* Go ahead and wire up outgoing edges and update PHIs for the duplicate | |
| 111 | 1280 block. */ |
| 1281 ssa_fix_duplicate_block_edges (rd, local_info); | |
| 0 | 1282 } |
| 1283 | |
| 1284 /* Keep walking the hash table. */ | |
| 1285 return 1; | |
| 1286 } | |
| 1287 | |
| 1288 /* We did not create any outgoing edges for the template block during | |
| 1289 block creation. This hash table traversal callback creates the | |
| 1290 outgoing edge for the template block. */ | |
| 1291 | |
| 111 | 1292 inline int |
| 1293 ssa_fixup_template_block (struct redirection_data **slot, | |
| 1294 ssa_local_info_t *local_info) | |
| 0 | 1295 { |
| 111 | 1296 struct redirection_data *rd = *slot; |
| 1297 | |
| 1298 /* If this is the template block halt the traversal after updating | |
| 1299 it appropriately. | |
| 1300 | |
| 1301 If we were threading through an joiner block, then we want | |
| 1302 to keep its control statement and redirect an outgoing edge. | |
| 1303 Else we want to remove the control statement & edges, then create | |
| 1304 a new outgoing edge. In both cases we may need to update PHIs. */ | |
| 1305 if (rd->dup_blocks[0] && rd->dup_blocks[0] == local_info->template_block) | |
| 0 | 1306 { |
| 111 | 1307 ssa_fix_duplicate_block_edges (rd, local_info); |
| 0 | 1308 return 0; |
| 1309 } | |
| 1310 | |
| 1311 return 1; | |
| 1312 } | |
| 1313 | |
| 1314 /* Hash table traversal callback to redirect each incoming edge | |
| 1315 associated with this hash table element to its new destination. */ | |
| 1316 | |
| 111 | 1317 int |
| 1318 ssa_redirect_edges (struct redirection_data **slot, | |
| 1319 ssa_local_info_t *local_info) | |
| 0 | 1320 { |
| 111 | 1321 struct redirection_data *rd = *slot; |
| 0 | 1322 struct el *next, *el; |
| 1323 | |
| 111 | 1324 /* Walk over all the incoming edges associated with this hash table |
| 1325 entry. */ | |
| 0 | 1326 for (el = rd->incoming_edges; el; el = next) |
| 1327 { | |
| 1328 edge e = el->e; | |
| 111 | 1329 vec<jump_thread_edge *> *path = THREAD_PATH (e); |
| 0 | 1330 |
| 1331 /* Go ahead and free this element from the list. Doing this now | |
| 1332 avoids the need for another list walk when we destroy the hash | |
| 1333 table. */ | |
| 1334 next = el->next; | |
| 1335 free (el); | |
| 1336 | |
| 1337 thread_stats.num_threaded_edges++; | |
| 1338 | |
| 111 | 1339 if (rd->dup_blocks[0]) |
| 0 | 1340 { |
| 1341 edge e2; | |
| 1342 | |
| 1343 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 1344 fprintf (dump_file, " Threaded jump %d --> %d to %d\n", | |
| 111 | 1345 e->src->index, e->dest->index, rd->dup_blocks[0]->index); |
| 1346 | |
| 1347 /* Redirect the incoming edge (possibly to the joiner block) to the | |
| 1348 appropriate duplicate block. */ | |
| 1349 e2 = redirect_edge_and_branch (e, rd->dup_blocks[0]); | |
| 0 | 1350 gcc_assert (e == e2); |
| 1351 flush_pending_stmts (e2); | |
| 1352 } | |
| 111 | 1353 |
| 1354 /* Go ahead and clear E->aux. It's not needed anymore and failure | |
| 1355 to clear it will cause all kinds of unpleasant problems later. */ | |
| 1356 delete_jump_thread_path (path); | |
| 1357 e->aux = NULL; | |
| 1358 | |
| 0 | 1359 } |
| 1360 | |
| 1361 /* Indicate that we actually threaded one or more jumps. */ | |
| 1362 if (rd->incoming_edges) | |
| 1363 local_info->jumps_threaded = true; | |
| 1364 | |
| 1365 return 1; | |
| 1366 } | |
| 1367 | |
| 1368 /* Return true if this block has no executable statements other than | |
| 1369 a simple ctrl flow instruction. When the number of outgoing edges | |
| 1370 is one, this is equivalent to a "forwarder" block. */ | |
| 1371 | |
| 1372 static bool | |
| 1373 redirection_block_p (basic_block bb) | |
| 1374 { | |
| 1375 gimple_stmt_iterator gsi; | |
| 1376 | |
| 1377 /* Advance to the first executable statement. */ | |
| 1378 gsi = gsi_start_bb (bb); | |
| 1379 while (!gsi_end_p (gsi) | |
| 111 | 1380 && (gimple_code (gsi_stmt (gsi)) == GIMPLE_LABEL |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1381 || is_gimple_debug (gsi_stmt (gsi)) |
| 111 | 1382 || gimple_nop_p (gsi_stmt (gsi)) |
| 1383 || gimple_clobber_p (gsi_stmt (gsi)))) | |
| 0 | 1384 gsi_next (&gsi); |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1385 |
| 0 | 1386 /* Check if this is an empty block. */ |
| 1387 if (gsi_end_p (gsi)) | |
| 1388 return true; | |
| 1389 | |
| 1390 /* Test that we've reached the terminating control statement. */ | |
| 1391 return gsi_stmt (gsi) | |
| 111 | 1392 && (gimple_code (gsi_stmt (gsi)) == GIMPLE_COND |
| 1393 || gimple_code (gsi_stmt (gsi)) == GIMPLE_GOTO | |
| 1394 || gimple_code (gsi_stmt (gsi)) == GIMPLE_SWITCH); | |
| 0 | 1395 } |
| 1396 | |
| 1397 /* BB is a block which ends with a COND_EXPR or SWITCH_EXPR and when BB | |
| 1398 is reached via one or more specific incoming edges, we know which | |
| 1399 outgoing edge from BB will be traversed. | |
| 1400 | |
| 1401 We want to redirect those incoming edges to the target of the | |
| 1402 appropriate outgoing edge. Doing so avoids a conditional branch | |
| 1403 and may expose new optimization opportunities. Note that we have | |
| 1404 to update dominator tree and SSA graph after such changes. | |
| 1405 | |
| 1406 The key to keeping the SSA graph update manageable is to duplicate | |
| 1407 the side effects occurring in BB so that those side effects still | |
| 1408 occur on the paths which bypass BB after redirecting edges. | |
| 1409 | |
| 1410 We accomplish this by creating duplicates of BB and arranging for | |
| 1411 the duplicates to unconditionally pass control to one specific | |
| 1412 successor of BB. We then revector the incoming edges into BB to | |
| 1413 the appropriate duplicate of BB. | |
| 1414 | |
| 1415 If NOLOOP_ONLY is true, we only perform the threading as long as it | |
| 111 | 1416 does not affect the structure of the loops in a nontrivial way. |
| 1417 | |
| 1418 If JOINERS is true, then thread through joiner blocks as well. */ | |
| 0 | 1419 |
| 1420 static bool | |
| 111 | 1421 thread_block_1 (basic_block bb, bool noloop_only, bool joiners) |
| 0 | 1422 { |
| 1423 /* E is an incoming edge into BB that we may or may not want to | |
| 1424 redirect to a duplicate of BB. */ | |
| 1425 edge e, e2; | |
| 1426 edge_iterator ei; | |
| 111 | 1427 ssa_local_info_t local_info; |
| 1428 | |
| 1429 local_info.duplicate_blocks = BITMAP_ALLOC (NULL); | |
| 1430 local_info.need_profile_correction = false; | |
| 0 | 1431 |
| 1432 /* To avoid scanning a linear array for the element we need we instead | |
| 1433 use a hash table. For normal code there should be no noticeable | |
| 1434 difference. However, if we have a block with a large number of | |
| 1435 incoming and outgoing edges such linear searches can get expensive. */ | |
| 111 | 1436 redirection_data |
| 1437 = new hash_table<struct redirection_data> (EDGE_COUNT (bb->succs)); | |
| 0 | 1438 |
| 1439 /* Record each unique threaded destination into a hash table for | |
| 1440 efficient lookups. */ | |
| 111 | 1441 edge last = NULL; |
| 0 | 1442 FOR_EACH_EDGE (e, ei, bb->preds) |
| 1443 { | |
| 111 | 1444 if (e->aux == NULL) |
| 1445 continue; | |
| 1446 | |
| 1447 vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
| 1448 | |
| 1449 if (((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK && !joiners) | |
| 1450 || ((*path)[1]->type == EDGE_COPY_SRC_BLOCK && joiners)) | |
| 1451 continue; | |
| 1452 | |
| 1453 e2 = path->last ()->e; | |
| 1454 if (!e2 || noloop_only) | |
| 1455 { | |
| 0 | 1456 /* If NOLOOP_ONLY is true, we only allow threading through the |
| 1457 header of a loop to exit edges. */ | |
| 111 | 1458 |
| 1459 /* One case occurs when there was loop header buried in a jump | |
| 1460 threading path that crosses loop boundaries. We do not try | |
| 1461 and thread this elsewhere, so just cancel the jump threading | |
| 1462 request by clearing the AUX field now. */ | |
| 1463 if (bb->loop_father != e2->src->loop_father | |
| 1464 && !loop_exit_edge_p (e2->src->loop_father, e2)) | |
| 1465 { | |
| 1466 /* Since this case is not handled by our special code | |
| 1467 to thread through a loop header, we must explicitly | |
| 1468 cancel the threading request here. */ | |
| 1469 delete_jump_thread_path (path); | |
| 1470 e->aux = NULL; | |
| 1471 continue; | |
| 1472 } | |
| 1473 | |
| 1474 /* Another case occurs when trying to thread through our | |
| 1475 own loop header, possibly from inside the loop. We will | |
| 1476 thread these later. */ | |
| 1477 unsigned int i; | |
| 1478 for (i = 1; i < path->length (); i++) | |
| 1479 { | |
| 1480 if ((*path)[i]->e->src == bb->loop_father->header | |
| 1481 && (!loop_exit_edge_p (bb->loop_father, e2) | |
| 1482 || (*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK)) | |
| 1483 break; | |
| 1484 } | |
| 1485 | |
| 1486 if (i != path->length ()) | |
| 1487 continue; | |
| 0 | 1488 } |
| 1489 | |
| 1490 /* Insert the outgoing edge into the hash table if it is not | |
| 1491 already in the hash table. */ | |
| 111 | 1492 lookup_redirection_data (e, INSERT); |
| 1493 | |
| 1494 /* When we have thread paths through a common joiner with different | |
| 1495 final destinations, then we may need corrections to deal with | |
| 1496 profile insanities. See the big comment before compute_path_counts. */ | |
| 1497 if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
| 1498 { | |
| 1499 if (!last) | |
| 1500 last = e2; | |
| 1501 else if (e2 != last) | |
| 1502 local_info.need_profile_correction = true; | |
| 1503 } | |
| 0 | 1504 } |
| 1505 | |
| 1506 /* We do not update dominance info. */ | |
| 1507 free_dominance_info (CDI_DOMINATORS); | |
| 1508 | |
| 111 | 1509 /* We know we only thread through the loop header to loop exits. |
| 1510 Let the basic block duplication hook know we are not creating | |
| 1511 a multiple entry loop. */ | |
| 1512 if (noloop_only | |
| 1513 && bb == bb->loop_father->header) | |
| 1514 set_loop_copy (bb->loop_father, loop_outer (bb->loop_father)); | |
| 1515 | |
| 0 | 1516 /* Now create duplicates of BB. |
| 1517 | |
| 1518 Note that for a block with a high outgoing degree we can waste | |
| 1519 a lot of time and memory creating and destroying useless edges. | |
| 1520 | |
| 1521 So we first duplicate BB and remove the control structure at the | |
| 1522 tail of the duplicate as well as all outgoing edges from the | |
| 1523 duplicate. We then use that duplicate block as a template for | |
| 1524 the rest of the duplicates. */ | |
| 1525 local_info.template_block = NULL; | |
| 1526 local_info.bb = bb; | |
| 1527 local_info.jumps_threaded = false; | |
| 111 | 1528 redirection_data->traverse <ssa_local_info_t *, ssa_create_duplicates> |
| 1529 (&local_info); | |
| 0 | 1530 |
| 1531 /* The template does not have an outgoing edge. Create that outgoing | |
| 1532 edge and update PHI nodes as the edge's target as necessary. | |
| 1533 | |
| 1534 We do this after creating all the duplicates to avoid creating | |
| 1535 unnecessary edges. */ | |
| 111 | 1536 redirection_data->traverse <ssa_local_info_t *, ssa_fixup_template_block> |
| 1537 (&local_info); | |
| 0 | 1538 |
| 1539 /* The hash table traversals above created the duplicate blocks (and the | |
| 1540 statements within the duplicate blocks). This loop creates PHI nodes for | |
| 1541 the duplicated blocks and redirects the incoming edges into BB to reach | |
| 1542 the duplicates of BB. */ | |
| 111 | 1543 redirection_data->traverse <ssa_local_info_t *, ssa_redirect_edges> |
| 1544 (&local_info); | |
| 0 | 1545 |
| 1546 /* Done with this block. Clear REDIRECTION_DATA. */ | |
| 111 | 1547 delete redirection_data; |
| 0 | 1548 redirection_data = NULL; |
| 1549 | |
| 111 | 1550 if (noloop_only |
| 1551 && bb == bb->loop_father->header) | |
| 1552 set_loop_copy (bb->loop_father, NULL); | |
| 1553 | |
| 1554 BITMAP_FREE (local_info.duplicate_blocks); | |
| 1555 local_info.duplicate_blocks = NULL; | |
| 1556 | |
| 0 | 1557 /* Indicate to our caller whether or not any jumps were threaded. */ |
| 1558 return local_info.jumps_threaded; | |
| 1559 } | |
| 1560 | |
| 111 | 1561 /* Wrapper for thread_block_1 so that we can first handle jump |
| 1562 thread paths which do not involve copying joiner blocks, then | |
| 1563 handle jump thread paths which have joiner blocks. | |
| 1564 | |
| 1565 By doing things this way we can be as aggressive as possible and | |
| 1566 not worry that copying a joiner block will create a jump threading | |
| 1567 opportunity. */ | |
| 1568 | |
| 1569 static bool | |
| 1570 thread_block (basic_block bb, bool noloop_only) | |
| 0 | 1571 { |
| 111 | 1572 bool retval; |
| 1573 retval = thread_block_1 (bb, noloop_only, false); | |
| 1574 retval |= thread_block_1 (bb, noloop_only, true); | |
| 1575 return retval; | |
| 0 | 1576 } |
| 1577 | |
| 1578 /* Callback for dfs_enumerate_from. Returns true if BB is different | |
| 1579 from STOP and DBDS_CE_STOP. */ | |
| 1580 | |
| 1581 static basic_block dbds_ce_stop; | |
| 1582 static bool | |
| 1583 dbds_continue_enumeration_p (const_basic_block bb, const void *stop) | |
| 1584 { | |
| 1585 return (bb != (const_basic_block) stop | |
| 1586 && bb != dbds_ce_stop); | |
| 1587 } | |
| 1588 | |
| 1589 /* Evaluates the dominance relationship of latch of the LOOP and BB, and | |
| 1590 returns the state. */ | |
| 1591 | |
| 1592 enum bb_dom_status | |
| 1593 determine_bb_domination_status (struct loop *loop, basic_block bb) | |
| 1594 { | |
| 1595 basic_block *bblocks; | |
| 1596 unsigned nblocks, i; | |
| 1597 bool bb_reachable = false; | |
| 1598 edge_iterator ei; | |
| 1599 edge e; | |
| 1600 | |
| 111 | 1601 /* This function assumes BB is a successor of LOOP->header. |
| 1602 If that is not the case return DOMST_NONDOMINATING which | |
| 1603 is always safe. */ | |
| 0 | 1604 { |
| 1605 bool ok = false; | |
| 1606 | |
| 1607 FOR_EACH_EDGE (e, ei, bb->preds) | |
| 1608 { | |
| 1609 if (e->src == loop->header) | |
| 1610 { | |
| 1611 ok = true; | |
| 1612 break; | |
| 1613 } | |
| 1614 } | |
| 1615 | |
| 111 | 1616 if (!ok) |
| 1617 return DOMST_NONDOMINATING; | |
| 0 | 1618 } |
| 1619 | |
| 1620 if (bb == loop->latch) | |
| 1621 return DOMST_DOMINATING; | |
| 1622 | |
| 1623 /* Check that BB dominates LOOP->latch, and that it is back-reachable | |
| 1624 from it. */ | |
| 1625 | |
| 1626 bblocks = XCNEWVEC (basic_block, loop->num_nodes); | |
| 1627 dbds_ce_stop = loop->header; | |
| 1628 nblocks = dfs_enumerate_from (loop->latch, 1, dbds_continue_enumeration_p, | |
| 1629 bblocks, loop->num_nodes, bb); | |
| 1630 for (i = 0; i < nblocks; i++) | |
| 1631 FOR_EACH_EDGE (e, ei, bblocks[i]->preds) | |
| 1632 { | |
| 1633 if (e->src == loop->header) | |
| 1634 { | |
| 1635 free (bblocks); | |
| 1636 return DOMST_NONDOMINATING; | |
| 1637 } | |
| 1638 if (e->src == bb) | |
| 1639 bb_reachable = true; | |
| 1640 } | |
| 1641 | |
| 1642 free (bblocks); | |
| 1643 return (bb_reachable ? DOMST_DOMINATING : DOMST_LOOP_BROKEN); | |
| 1644 } | |
| 1645 | |
| 1646 /* Thread jumps through the header of LOOP. Returns true if cfg changes. | |
| 1647 If MAY_PEEL_LOOP_HEADERS is false, we avoid threading from entry edges | |
| 1648 to the inside of the loop. */ | |
| 1649 | |
| 1650 static bool | |
| 1651 thread_through_loop_header (struct loop *loop, bool may_peel_loop_headers) | |
| 1652 { | |
| 1653 basic_block header = loop->header; | |
| 1654 edge e, tgt_edge, latch = loop_latch_edge (loop); | |
| 1655 edge_iterator ei; | |
| 1656 basic_block tgt_bb, atgt_bb; | |
| 1657 enum bb_dom_status domst; | |
| 1658 | |
| 1659 /* We have already threaded through headers to exits, so all the threading | |
| 1660 requests now are to the inside of the loop. We need to avoid creating | |
| 1661 irreducible regions (i.e., loops with more than one entry block), and | |
| 1662 also loop with several latch edges, or new subloops of the loop (although | |
| 1663 there are cases where it might be appropriate, it is difficult to decide, | |
| 1664 and doing it wrongly may confuse other optimizers). | |
| 1665 | |
| 1666 We could handle more general cases here. However, the intention is to | |
| 1667 preserve some information about the loop, which is impossible if its | |
| 1668 structure changes significantly, in a way that is not well understood. | |
| 1669 Thus we only handle few important special cases, in which also updating | |
| 1670 of the loop-carried information should be feasible: | |
| 1671 | |
| 1672 1) Propagation of latch edge to a block that dominates the latch block | |
| 1673 of a loop. This aims to handle the following idiom: | |
| 1674 | |
| 1675 first = 1; | |
| 1676 while (1) | |
| 1677 { | |
| 1678 if (first) | |
| 1679 initialize; | |
| 1680 first = 0; | |
| 1681 body; | |
| 1682 } | |
| 1683 | |
| 1684 After threading the latch edge, this becomes | |
| 1685 | |
| 1686 first = 1; | |
| 1687 if (first) | |
| 1688 initialize; | |
| 1689 while (1) | |
| 1690 { | |
| 1691 first = 0; | |
| 1692 body; | |
| 1693 } | |
| 1694 | |
| 1695 The original header of the loop is moved out of it, and we may thread | |
| 1696 the remaining edges through it without further constraints. | |
| 1697 | |
| 1698 2) All entry edges are propagated to a single basic block that dominates | |
| 1699 the latch block of the loop. This aims to handle the following idiom | |
| 1700 (normally created for "for" loops): | |
| 1701 | |
| 1702 i = 0; | |
| 1703 while (1) | |
| 1704 { | |
| 1705 if (i >= 100) | |
| 1706 break; | |
| 1707 body; | |
| 1708 i++; | |
| 1709 } | |
| 1710 | |
| 1711 This becomes | |
| 1712 | |
| 1713 i = 0; | |
| 1714 while (1) | |
| 1715 { | |
| 1716 body; | |
| 1717 i++; | |
| 1718 if (i >= 100) | |
| 1719 break; | |
| 1720 } | |
| 1721 */ | |
| 1722 | |
| 1723 /* Threading through the header won't improve the code if the header has just | |
| 1724 one successor. */ | |
| 1725 if (single_succ_p (header)) | |
| 1726 goto fail; | |
| 1727 | |
| 111 | 1728 if (!may_peel_loop_headers && !redirection_block_p (loop->header)) |
| 0 | 1729 goto fail; |
| 1730 else | |
| 1731 { | |
| 1732 tgt_bb = NULL; | |
| 1733 tgt_edge = NULL; | |
| 1734 FOR_EACH_EDGE (e, ei, header->preds) | |
| 1735 { | |
| 1736 if (!e->aux) | |
| 1737 { | |
| 1738 if (e == latch) | |
| 1739 continue; | |
| 1740 | |
| 1741 /* If latch is not threaded, and there is a header | |
| 1742 edge that is not threaded, we would create loop | |
| 1743 with multiple entries. */ | |
| 1744 goto fail; | |
| 1745 } | |
| 1746 | |
| 111 | 1747 vec<jump_thread_edge *> *path = THREAD_PATH (e); |
| 1748 | |
| 1749 if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
| 1750 goto fail; | |
| 1751 tgt_edge = (*path)[1]->e; | |
| 0 | 1752 atgt_bb = tgt_edge->dest; |
| 1753 if (!tgt_bb) | |
| 1754 tgt_bb = atgt_bb; | |
| 1755 /* Two targets of threading would make us create loop | |
| 1756 with multiple entries. */ | |
| 1757 else if (tgt_bb != atgt_bb) | |
| 1758 goto fail; | |
| 1759 } | |
| 1760 | |
| 1761 if (!tgt_bb) | |
| 1762 { | |
| 1763 /* There are no threading requests. */ | |
| 1764 return false; | |
| 1765 } | |
| 1766 | |
| 1767 /* Redirecting to empty loop latch is useless. */ | |
| 1768 if (tgt_bb == loop->latch | |
| 1769 && empty_block_p (loop->latch)) | |
| 1770 goto fail; | |
| 1771 } | |
| 1772 | |
| 1773 /* The target block must dominate the loop latch, otherwise we would be | |
| 1774 creating a subloop. */ | |
| 1775 domst = determine_bb_domination_status (loop, tgt_bb); | |
| 1776 if (domst == DOMST_NONDOMINATING) | |
| 1777 goto fail; | |
| 1778 if (domst == DOMST_LOOP_BROKEN) | |
| 1779 { | |
| 1780 /* If the loop ceased to exist, mark it as such, and thread through its | |
| 1781 original header. */ | |
| 111 | 1782 mark_loop_for_removal (loop); |
| 0 | 1783 return thread_block (header, false); |
| 1784 } | |
| 1785 | |
| 1786 if (tgt_bb->loop_father->header == tgt_bb) | |
| 1787 { | |
| 1788 /* If the target of the threading is a header of a subloop, we need | |
| 1789 to create a preheader for it, so that the headers of the two loops | |
| 1790 do not merge. */ | |
| 1791 if (EDGE_COUNT (tgt_bb->preds) > 2) | |
| 1792 { | |
| 1793 tgt_bb = create_preheader (tgt_bb->loop_father, 0); | |
| 1794 gcc_assert (tgt_bb != NULL); | |
| 1795 } | |
| 1796 else | |
| 1797 tgt_bb = split_edge (tgt_edge); | |
| 1798 } | |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1799 |
| 111 | 1800 basic_block new_preheader; |
| 1801 | |
| 1802 /* Now consider the case entry edges are redirected to the new entry | |
| 1803 block. Remember one entry edge, so that we can find the new | |
| 1804 preheader (its destination after threading). */ | |
| 1805 FOR_EACH_EDGE (e, ei, header->preds) | |
| 0 | 1806 { |
| 111 | 1807 if (e->aux) |
| 1808 break; | |
| 0 | 1809 } |
|
55
77e2b8dfacca
update it from 4.4.3 to 4.5.0
ryoma <e075725@ie.u-ryukyu.ac.jp>
parents:
0
diff
changeset
|
1810 |
| 111 | 1811 /* The duplicate of the header is the new preheader of the loop. Ensure |
| 1812 that it is placed correctly in the loop hierarchy. */ | |
| 1813 set_loop_copy (loop, loop_outer (loop)); | |
| 1814 | |
| 1815 thread_block (header, false); | |
| 1816 set_loop_copy (loop, NULL); | |
| 1817 new_preheader = e->dest; | |
| 1818 | |
| 1819 /* Create the new latch block. This is always necessary, as the latch | |
| 1820 must have only a single successor, but the original header had at | |
| 1821 least two successors. */ | |
| 1822 loop->latch = NULL; | |
| 1823 mfb_kj_edge = single_succ_edge (new_preheader); | |
| 1824 loop->header = mfb_kj_edge->dest; | |
| 1825 latch = make_forwarder_block (tgt_bb, mfb_keep_just, NULL); | |
| 1826 loop->header = latch->dest; | |
| 1827 loop->latch = latch->src; | |
| 0 | 1828 return true; |
| 1829 | |
| 1830 fail: | |
| 1831 /* We failed to thread anything. Cancel the requests. */ | |
| 1832 FOR_EACH_EDGE (e, ei, header->preds) | |
| 1833 { | |
| 111 | 1834 vec<jump_thread_edge *> *path = THREAD_PATH (e); |
| 1835 | |
| 1836 if (path) | |
| 1837 { | |
| 1838 delete_jump_thread_path (path); | |
| 1839 e->aux = NULL; | |
| 1840 } | |
| 0 | 1841 } |
| 1842 return false; | |
| 1843 } | |
| 1844 | |
| 111 | 1845 /* E1 and E2 are edges into the same basic block. Return TRUE if the |
| 1846 PHI arguments associated with those edges are equal or there are no | |
| 1847 PHI arguments, otherwise return FALSE. */ | |
| 1848 | |
| 1849 static bool | |
| 1850 phi_args_equal_on_edges (edge e1, edge e2) | |
| 1851 { | |
| 1852 gphi_iterator gsi; | |
| 1853 int indx1 = e1->dest_idx; | |
| 1854 int indx2 = e2->dest_idx; | |
| 1855 | |
| 1856 for (gsi = gsi_start_phis (e1->dest); !gsi_end_p (gsi); gsi_next (&gsi)) | |
| 1857 { | |
| 1858 gphi *phi = gsi.phi (); | |
| 1859 | |
| 1860 if (!operand_equal_p (gimple_phi_arg_def (phi, indx1), | |
| 1861 gimple_phi_arg_def (phi, indx2), 0)) | |
| 1862 return false; | |
| 1863 } | |
| 1864 return true; | |
| 1865 } | |
| 1866 | |
| 0 | 1867 /* Walk through the registered jump threads and convert them into a |
| 1868 form convenient for this pass. | |
| 1869 | |
| 1870 Any block which has incoming edges threaded to outgoing edges | |
| 1871 will have its entry in THREADED_BLOCK set. | |
| 1872 | |
| 1873 Any threaded edge will have its new outgoing edge stored in the | |
| 1874 original edge's AUX field. | |
| 1875 | |
| 1876 This form avoids the need to walk all the edges in the CFG to | |
| 1877 discover blocks which need processing and avoids unnecessary | |
| 1878 hash table lookups to map from threaded edge to new target. */ | |
| 1879 | |
| 1880 static void | |
| 1881 mark_threaded_blocks (bitmap threaded_blocks) | |
| 1882 { | |
| 1883 unsigned int i; | |
| 1884 bitmap_iterator bi; | |
| 111 | 1885 auto_bitmap tmp; |
| 0 | 1886 basic_block bb; |
| 1887 edge e; | |
| 1888 edge_iterator ei; | |
| 1889 | |
| 111 | 1890 /* It is possible to have jump threads in which one is a subpath |
| 1891 of the other. ie, (A, B), (B, C), (C, D) where B is a joiner | |
| 1892 block and (B, C), (C, D) where no joiner block exists. | |
| 1893 | |
| 1894 When this occurs ignore the jump thread request with the joiner | |
| 1895 block. It's totally subsumed by the simpler jump thread request. | |
| 1896 | |
| 1897 This results in less block copying, simpler CFGs. More importantly, | |
| 1898 when we duplicate the joiner block, B, in this case we will create | |
| 1899 a new threading opportunity that we wouldn't be able to optimize | |
| 1900 until the next jump threading iteration. | |
| 1901 | |
| 1902 So first convert the jump thread requests which do not require a | |
| 1903 joiner block. */ | |
| 1904 for (i = 0; i < paths.length (); i++) | |
| 1905 { | |
| 1906 vec<jump_thread_edge *> *path = paths[i]; | |
| 1907 | |
| 1908 if ((*path)[1]->type != EDGE_COPY_SRC_JOINER_BLOCK) | |
| 1909 { | |
| 1910 edge e = (*path)[0]->e; | |
| 1911 e->aux = (void *)path; | |
| 1912 bitmap_set_bit (tmp, e->dest->index); | |
| 1913 } | |
| 1914 } | |
| 1915 | |
| 1916 /* Now iterate again, converting cases where we want to thread | |
| 1917 through a joiner block, but only if no other edge on the path | |
| 1918 already has a jump thread attached to it. We do this in two passes, | |
| 1919 to avoid situations where the order in the paths vec can hide overlapping | |
| 1920 threads (the path is recorded on the incoming edge, so we would miss | |
| 1921 cases where the second path starts at a downstream edge on the same | |
| 1922 path). First record all joiner paths, deleting any in the unexpected | |
| 1923 case where there is already a path for that incoming edge. */ | |
| 1924 for (i = 0; i < paths.length ();) | |
| 0 | 1925 { |
| 111 | 1926 vec<jump_thread_edge *> *path = paths[i]; |
| 1927 | |
| 1928 if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK) | |
| 1929 { | |
| 1930 /* Attach the path to the starting edge if none is yet recorded. */ | |
| 1931 if ((*path)[0]->e->aux == NULL) | |
| 1932 { | |
| 1933 (*path)[0]->e->aux = path; | |
| 1934 i++; | |
| 1935 } | |
| 1936 else | |
| 1937 { | |
| 1938 paths.unordered_remove (i); | |
| 1939 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 1940 dump_jump_thread_path (dump_file, *path, false); | |
| 1941 delete_jump_thread_path (path); | |
| 1942 } | |
| 1943 } | |
| 1944 else | |
| 1945 { | |
| 1946 i++; | |
| 1947 } | |
| 1948 } | |
| 1949 | |
| 1950 /* Second, look for paths that have any other jump thread attached to | |
| 1951 them, and either finish converting them or cancel them. */ | |
| 1952 for (i = 0; i < paths.length ();) | |
| 1953 { | |
| 1954 vec<jump_thread_edge *> *path = paths[i]; | |
| 1955 edge e = (*path)[0]->e; | |
| 1956 | |
| 1957 if ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK && e->aux == path) | |
| 1958 { | |
| 1959 unsigned int j; | |
| 1960 for (j = 1; j < path->length (); j++) | |
| 1961 if ((*path)[j]->e->aux != NULL) | |
| 1962 break; | |
| 1963 | |
| 1964 /* If we iterated through the entire path without exiting the loop, | |
| 1965 then we are good to go, record it. */ | |
| 1966 if (j == path->length ()) | |
| 1967 { | |
| 1968 bitmap_set_bit (tmp, e->dest->index); | |
| 1969 i++; | |
| 1970 } | |
| 1971 else | |
| 1972 { | |
| 1973 e->aux = NULL; | |
| 1974 paths.unordered_remove (i); | |
| 1975 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 1976 dump_jump_thread_path (dump_file, *path, false); | |
| 1977 delete_jump_thread_path (path); | |
| 1978 } | |
| 1979 } | |
| 1980 else | |
| 1981 { | |
| 1982 i++; | |
| 1983 } | |
| 0 | 1984 } |
| 1985 | |
| 1986 /* If optimizing for size, only thread through block if we don't have | |
| 1987 to duplicate it or it's an otherwise empty redirection block. */ | |
| 1988 if (optimize_function_for_size_p (cfun)) | |
| 1989 { | |
| 1990 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi) | |
| 1991 { | |
| 111 | 1992 bb = BASIC_BLOCK_FOR_FN (cfun, i); |
| 0 | 1993 if (EDGE_COUNT (bb->preds) > 1 |
| 1994 && !redirection_block_p (bb)) | |
| 1995 { | |
| 1996 FOR_EACH_EDGE (e, ei, bb->preds) | |
| 111 | 1997 { |
| 1998 if (e->aux) | |
| 1999 { | |
| 2000 vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
| 2001 delete_jump_thread_path (path); | |
| 0 | 2002 e->aux = NULL; |
| 111 | 2003 } |
| 2004 } | |
| 0 | 2005 } |
| 2006 else | |
| 2007 bitmap_set_bit (threaded_blocks, i); | |
| 2008 } | |
| 2009 } | |
| 2010 else | |
| 2011 bitmap_copy (threaded_blocks, tmp); | |
| 2012 | |
| 111 | 2013 /* If we have a joiner block (J) which has two successors S1 and S2 and |
| 2014 we are threading though S1 and the final destination of the thread | |
| 2015 is S2, then we must verify that any PHI nodes in S2 have the same | |
| 2016 PHI arguments for the edge J->S2 and J->S1->...->S2. | |
| 2017 | |
| 2018 We used to detect this prior to registering the jump thread, but | |
| 2019 that prohibits propagation of edge equivalences into non-dominated | |
| 2020 PHI nodes as the equivalency test might occur before propagation. | |
| 2021 | |
| 2022 This must also occur after we truncate any jump threading paths | |
| 2023 as this scenario may only show up after truncation. | |
| 2024 | |
| 2025 This works for now, but will need improvement as part of the FSA | |
| 2026 optimization. | |
| 2027 | |
| 2028 Note since we've moved the thread request data to the edges, | |
| 2029 we have to iterate on those rather than the threaded_edges vector. */ | |
| 2030 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi) | |
| 2031 { | |
| 2032 bb = BASIC_BLOCK_FOR_FN (cfun, i); | |
| 2033 FOR_EACH_EDGE (e, ei, bb->preds) | |
| 2034 { | |
| 2035 if (e->aux) | |
| 2036 { | |
| 2037 vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
| 2038 bool have_joiner = ((*path)[1]->type == EDGE_COPY_SRC_JOINER_BLOCK); | |
| 2039 | |
| 2040 if (have_joiner) | |
| 2041 { | |
| 2042 basic_block joiner = e->dest; | |
| 2043 edge final_edge = path->last ()->e; | |
| 2044 basic_block final_dest = final_edge->dest; | |
| 2045 edge e2 = find_edge (joiner, final_dest); | |
| 2046 | |
| 2047 if (e2 && !phi_args_equal_on_edges (e2, final_edge)) | |
| 2048 { | |
| 2049 delete_jump_thread_path (path); | |
| 2050 e->aux = NULL; | |
| 2051 } | |
| 2052 } | |
| 2053 } | |
| 2054 } | |
| 2055 } | |
| 2056 | |
| 2057 /* Look for jump threading paths which cross multiple loop headers. | |
| 2058 | |
| 2059 The code to thread through loop headers will change the CFG in ways | |
| 2060 that invalidate the cached loop iteration information. So we must | |
| 2061 detect that case and wipe the cached information. */ | |
| 2062 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi) | |
| 2063 { | |
| 2064 basic_block bb = BASIC_BLOCK_FOR_FN (cfun, i); | |
| 2065 FOR_EACH_EDGE (e, ei, bb->preds) | |
| 2066 { | |
| 2067 if (e->aux) | |
| 2068 { | |
| 2069 vec<jump_thread_edge *> *path = THREAD_PATH (e); | |
| 2070 | |
| 2071 for (unsigned int i = 0, crossed_headers = 0; | |
| 2072 i < path->length (); | |
| 2073 i++) | |
| 2074 { | |
| 2075 basic_block dest = (*path)[i]->e->dest; | |
| 2076 basic_block src = (*path)[i]->e->src; | |
| 2077 /* If we enter a loop. */ | |
| 2078 if (flow_loop_nested_p (src->loop_father, dest->loop_father)) | |
| 2079 ++crossed_headers; | |
| 2080 /* If we step from a block outside an irreducible region | |
| 2081 to a block inside an irreducible region, then we have | |
| 2082 crossed into a loop. */ | |
| 2083 else if (! (src->flags & BB_IRREDUCIBLE_LOOP) | |
| 2084 && (dest->flags & BB_IRREDUCIBLE_LOOP)) | |
| 2085 ++crossed_headers; | |
| 2086 if (crossed_headers > 1) | |
| 2087 { | |
| 2088 vect_free_loop_info_assumptions | |
| 2089 ((*path)[path->length () - 1]->e->dest->loop_father); | |
| 2090 break; | |
| 2091 } | |
| 2092 } | |
| 2093 } | |
| 2094 } | |
| 2095 } | |
| 2096 } | |
| 2097 | |
| 2098 | |
| 2099 /* Verify that the REGION is a valid jump thread. A jump thread is a special | |
| 2100 case of SEME Single Entry Multiple Exits region in which all nodes in the | |
| 2101 REGION have exactly one incoming edge. The only exception is the first block | |
| 2102 that may not have been connected to the rest of the cfg yet. */ | |
| 2103 | |
| 2104 DEBUG_FUNCTION void | |
| 2105 verify_jump_thread (basic_block *region, unsigned n_region) | |
| 2106 { | |
| 2107 for (unsigned i = 0; i < n_region; i++) | |
| 2108 gcc_assert (EDGE_COUNT (region[i]->preds) <= 1); | |
| 2109 } | |
| 2110 | |
| 2111 /* Return true when BB is one of the first N items in BBS. */ | |
| 2112 | |
| 2113 static inline bool | |
| 2114 bb_in_bbs (basic_block bb, basic_block *bbs, int n) | |
| 2115 { | |
| 2116 for (int i = 0; i < n; i++) | |
| 2117 if (bb == bbs[i]) | |
| 2118 return true; | |
| 2119 | |
| 2120 return false; | |
| 0 | 2121 } |
| 2122 | |
| 111 | 2123 /* Duplicates a jump-thread path of N_REGION basic blocks. |
| 2124 The ENTRY edge is redirected to the duplicate of the region. | |
| 2125 | |
| 2126 Remove the last conditional statement in the last basic block in the REGION, | |
| 2127 and create a single fallthru edge pointing to the same destination as the | |
| 2128 EXIT edge. | |
| 2129 | |
| 2130 Returns false if it is unable to copy the region, true otherwise. */ | |
| 2131 | |
| 2132 static bool | |
| 2133 duplicate_thread_path (edge entry, edge exit, basic_block *region, | |
| 2134 unsigned n_region) | |
| 2135 { | |
| 2136 unsigned i; | |
| 2137 struct loop *loop = entry->dest->loop_father; | |
| 2138 edge exit_copy; | |
| 2139 edge redirected; | |
| 2140 int curr_freq; | |
| 2141 profile_count curr_count; | |
| 2142 | |
| 2143 if (!can_copy_bbs_p (region, n_region)) | |
| 2144 return false; | |
| 2145 | |
| 2146 /* Some sanity checking. Note that we do not check for all possible | |
| 2147 missuses of the functions. I.e. if you ask to copy something weird, | |
| 2148 it will work, but the state of structures probably will not be | |
| 2149 correct. */ | |
| 2150 for (i = 0; i < n_region; i++) | |
| 2151 { | |
| 2152 /* We do not handle subloops, i.e. all the blocks must belong to the | |
| 2153 same loop. */ | |
| 2154 if (region[i]->loop_father != loop) | |
| 2155 return false; | |
| 2156 } | |
| 2157 | |
| 2158 initialize_original_copy_tables (); | |
| 2159 | |
| 2160 set_loop_copy (loop, loop); | |
| 2161 | |
| 2162 basic_block *region_copy = XNEWVEC (basic_block, n_region); | |
| 2163 copy_bbs (region, n_region, region_copy, &exit, 1, &exit_copy, loop, | |
| 2164 split_edge_bb_loc (entry), false); | |
| 2165 | |
| 2166 /* Fix up: copy_bbs redirects all edges pointing to copied blocks. The | |
| 2167 following code ensures that all the edges exiting the jump-thread path are | |
| 2168 redirected back to the original code: these edges are exceptions | |
| 2169 invalidating the property that is propagated by executing all the blocks of | |
| 2170 the jump-thread path in order. */ | |
| 2171 | |
| 2172 curr_count = entry->count (); | |
| 2173 curr_freq = EDGE_FREQUENCY (entry); | |
| 2174 | |
| 2175 for (i = 0; i < n_region; i++) | |
| 2176 { | |
| 2177 edge e; | |
| 2178 edge_iterator ei; | |
| 2179 basic_block bb = region_copy[i]; | |
| 2180 | |
| 2181 /* Watch inconsistent profile. */ | |
| 2182 if (curr_count > region[i]->count) | |
| 2183 curr_count = region[i]->count; | |
| 2184 if (curr_freq > region[i]->frequency) | |
| 2185 curr_freq = region[i]->frequency; | |
| 2186 /* Scale current BB. */ | |
| 2187 if (region[i]->count > 0 && curr_count.initialized_p ()) | |
| 2188 { | |
| 2189 /* In the middle of the path we only scale the frequencies. | |
| 2190 In last BB we need to update probabilities of outgoing edges | |
| 2191 because we know which one is taken at the threaded path. */ | |
| 2192 if (i + 1 != n_region) | |
| 2193 scale_bbs_frequencies_profile_count (region + i, 1, | |
| 2194 region[i]->count - curr_count, | |
| 2195 region[i]->count); | |
| 2196 else | |
| 2197 update_bb_profile_for_threading (region[i], | |
| 2198 curr_freq, curr_count, | |
| 2199 exit); | |
| 2200 scale_bbs_frequencies_profile_count (region_copy + i, 1, curr_count, | |
| 2201 region_copy[i]->count); | |
| 2202 } | |
| 2203 else if (region[i]->frequency) | |
| 2204 { | |
| 2205 if (i + 1 != n_region) | |
| 2206 scale_bbs_frequencies_int (region + i, 1, | |
| 2207 region[i]->frequency - curr_freq, | |
| 2208 region[i]->frequency); | |
| 2209 else | |
| 2210 update_bb_profile_for_threading (region[i], | |
| 2211 curr_freq, curr_count, | |
| 2212 exit); | |
| 2213 scale_bbs_frequencies_int (region_copy + i, 1, curr_freq, | |
| 2214 region_copy[i]->frequency); | |
| 2215 } | |
| 2216 | |
| 2217 if (single_succ_p (bb)) | |
| 2218 { | |
| 2219 /* Make sure the successor is the next node in the path. */ | |
| 2220 gcc_assert (i + 1 == n_region | |
| 2221 || region_copy[i + 1] == single_succ_edge (bb)->dest); | |
| 2222 if (i + 1 != n_region) | |
| 2223 { | |
| 2224 curr_freq = EDGE_FREQUENCY (single_succ_edge (bb)); | |
| 2225 curr_count = single_succ_edge (bb)->count (); | |
| 2226 } | |
| 2227 continue; | |
| 2228 } | |
| 2229 | |
| 2230 /* Special case the last block on the path: make sure that it does not | |
| 2231 jump back on the copied path, including back to itself. */ | |
| 2232 if (i + 1 == n_region) | |
| 2233 { | |
| 2234 FOR_EACH_EDGE (e, ei, bb->succs) | |
| 2235 if (bb_in_bbs (e->dest, region_copy, n_region)) | |
| 2236 { | |
| 2237 basic_block orig = get_bb_original (e->dest); | |
| 2238 if (orig) | |
| 2239 redirect_edge_and_branch_force (e, orig); | |
| 2240 } | |
| 2241 continue; | |
| 2242 } | |
| 2243 | |
| 2244 /* Redirect all other edges jumping to non-adjacent blocks back to the | |
| 2245 original code. */ | |
| 2246 FOR_EACH_EDGE (e, ei, bb->succs) | |
| 2247 if (region_copy[i + 1] != e->dest) | |
| 2248 { | |
| 2249 basic_block orig = get_bb_original (e->dest); | |
| 2250 if (orig) | |
| 2251 redirect_edge_and_branch_force (e, orig); | |
| 2252 } | |
| 2253 else | |
| 2254 { | |
| 2255 curr_freq = EDGE_FREQUENCY (e); | |
| 2256 curr_count = e->count (); | |
| 2257 } | |
| 2258 } | |
| 2259 | |
| 2260 | |
| 2261 if (flag_checking) | |
| 2262 verify_jump_thread (region_copy, n_region); | |
| 2263 | |
| 2264 /* Remove the last branch in the jump thread path. */ | |
| 2265 remove_ctrl_stmt_and_useless_edges (region_copy[n_region - 1], exit->dest); | |
| 2266 | |
| 2267 /* And fixup the flags on the single remaining edge. */ | |
| 2268 edge fix_e = find_edge (region_copy[n_region - 1], exit->dest); | |
| 2269 fix_e->flags &= ~(EDGE_TRUE_VALUE | EDGE_FALSE_VALUE | EDGE_ABNORMAL); | |
| 2270 fix_e->flags |= EDGE_FALLTHRU; | |
| 2271 | |
| 2272 edge e = make_edge (region_copy[n_region - 1], exit->dest, EDGE_FALLTHRU); | |
| 2273 | |
| 2274 if (e) | |
| 2275 { | |
| 2276 rescan_loop_exit (e, true, false); | |
| 2277 e->probability = profile_probability::always (); | |
| 2278 } | |
| 2279 | |
| 2280 /* Redirect the entry and add the phi node arguments. */ | |
| 2281 if (entry->dest == loop->header) | |
| 2282 mark_loop_for_removal (loop); | |
| 2283 redirected = redirect_edge_and_branch (entry, get_bb_copy (entry->dest)); | |
| 2284 gcc_assert (redirected != NULL); | |
| 2285 flush_pending_stmts (entry); | |
| 2286 | |
| 2287 /* Add the other PHI node arguments. */ | |
| 2288 add_phi_args_after_copy (region_copy, n_region, NULL); | |
| 2289 | |
| 2290 free (region_copy); | |
| 2291 | |
| 2292 free_original_copy_tables (); | |
| 2293 return true; | |
| 2294 } | |
| 2295 | |
| 2296 /* Return true when PATH is a valid jump-thread path. */ | |
| 2297 | |
| 2298 static bool | |
| 2299 valid_jump_thread_path (vec<jump_thread_edge *> *path) | |
| 2300 { | |
| 2301 unsigned len = path->length (); | |
| 2302 | |
| 2303 /* Check that the path is connected. */ | |
| 2304 for (unsigned int j = 0; j < len - 1; j++) | |
| 2305 { | |
| 2306 edge e = (*path)[j]->e; | |
| 2307 if (e->dest != (*path)[j+1]->e->src) | |
| 2308 return false; | |
| 2309 } | |
| 2310 return true; | |
| 2311 } | |
| 2312 | |
| 2313 /* Remove any queued jump threads that include edge E. | |
| 2314 | |
| 2315 We don't actually remove them here, just record the edges into ax | |
| 2316 hash table. That way we can do the search once per iteration of | |
| 2317 DOM/VRP rather than for every case where DOM optimizes away a COND_EXPR. */ | |
| 2318 | |
| 2319 void | |
| 2320 remove_jump_threads_including (edge_def *e) | |
| 2321 { | |
| 2322 if (!paths.exists ()) | |
| 2323 return; | |
| 2324 | |
| 2325 if (!removed_edges) | |
| 2326 removed_edges = new hash_table<struct removed_edges> (17); | |
| 2327 | |
| 2328 edge *slot = removed_edges->find_slot (e, INSERT); | |
| 2329 *slot = e; | |
| 2330 } | |
| 0 | 2331 |
| 2332 /* Walk through all blocks and thread incoming edges to the appropriate | |
| 2333 outgoing edge for each edge pair recorded in THREADED_EDGES. | |
| 2334 | |
| 2335 It is the caller's responsibility to fix the dominance information | |
| 2336 and rewrite duplicated SSA_NAMEs back into SSA form. | |
| 2337 | |
| 2338 If MAY_PEEL_LOOP_HEADERS is false, we avoid threading edges through | |
| 2339 loop headers if it does not simplify the loop. | |
| 2340 | |
| 2341 Returns true if one or more edges were threaded, false otherwise. */ | |
| 2342 | |
| 2343 bool | |
| 2344 thread_through_all_blocks (bool may_peel_loop_headers) | |
| 2345 { | |
| 2346 bool retval = false; | |
| 2347 unsigned int i; | |
| 2348 bitmap_iterator bi; | |
| 2349 struct loop *loop; | |
| 111 | 2350 auto_bitmap threaded_blocks; |
| 2351 | |
| 2352 if (!paths.exists ()) | |
| 2353 { | |
| 2354 retval = false; | |
| 2355 goto out; | |
| 2356 } | |
| 2357 | |
| 0 | 2358 memset (&thread_stats, 0, sizeof (thread_stats)); |
| 2359 | |
| 111 | 2360 /* Remove any paths that referenced removed edges. */ |
| 2361 if (removed_edges) | |
| 2362 for (i = 0; i < paths.length (); ) | |
| 2363 { | |
| 2364 unsigned int j; | |
| 2365 vec<jump_thread_edge *> *path = paths[i]; | |
| 2366 | |
| 2367 for (j = 0; j < path->length (); j++) | |
| 2368 { | |
| 2369 edge e = (*path)[j]->e; | |
| 2370 if (removed_edges->find_slot (e, NO_INSERT)) | |
| 2371 break; | |
| 2372 } | |
| 2373 | |
| 2374 if (j != path->length ()) | |
| 2375 { | |
| 2376 delete_jump_thread_path (path); | |
| 2377 paths.unordered_remove (i); | |
| 2378 continue; | |
| 2379 } | |
| 2380 i++; | |
| 2381 } | |
| 2382 | |
| 2383 /* Jump-thread all FSM threads before other jump-threads. */ | |
| 2384 for (i = 0; i < paths.length ();) | |
| 2385 { | |
| 2386 vec<jump_thread_edge *> *path = paths[i]; | |
| 2387 edge entry = (*path)[0]->e; | |
| 2388 | |
| 2389 /* Only code-generate FSM jump-threads in this loop. */ | |
| 2390 if ((*path)[0]->type != EDGE_FSM_THREAD) | |
| 2391 { | |
| 2392 i++; | |
| 2393 continue; | |
| 2394 } | |
| 2395 | |
| 2396 /* Do not jump-thread twice from the same block. */ | |
| 2397 if (bitmap_bit_p (threaded_blocks, entry->src->index) | |
| 2398 /* We may not want to realize this jump thread path | |
| 2399 for various reasons. So check it first. */ | |
| 2400 || !valid_jump_thread_path (path)) | |
| 2401 { | |
| 2402 /* Remove invalid FSM jump-thread paths. */ | |
| 2403 delete_jump_thread_path (path); | |
| 2404 paths.unordered_remove (i); | |
| 2405 continue; | |
| 2406 } | |
| 2407 | |
| 2408 unsigned len = path->length (); | |
| 2409 edge exit = (*path)[len - 1]->e; | |
| 2410 basic_block *region = XNEWVEC (basic_block, len - 1); | |
| 2411 | |
| 2412 for (unsigned int j = 0; j < len - 1; j++) | |
| 2413 region[j] = (*path)[j]->e->dest; | |
| 2414 | |
| 2415 if (duplicate_thread_path (entry, exit, region, len - 1)) | |
| 2416 { | |
| 2417 /* We do not update dominance info. */ | |
| 2418 free_dominance_info (CDI_DOMINATORS); | |
| 2419 bitmap_set_bit (threaded_blocks, entry->src->index); | |
| 2420 retval = true; | |
| 2421 thread_stats.num_threaded_edges++; | |
| 2422 } | |
| 2423 | |
| 2424 delete_jump_thread_path (path); | |
| 2425 paths.unordered_remove (i); | |
| 2426 free (region); | |
| 2427 } | |
| 2428 | |
| 2429 /* Remove from PATHS all the jump-threads starting with an edge already | |
| 2430 jump-threaded. */ | |
| 2431 for (i = 0; i < paths.length ();) | |
| 2432 { | |
| 2433 vec<jump_thread_edge *> *path = paths[i]; | |
| 2434 edge entry = (*path)[0]->e; | |
| 2435 | |
| 2436 /* Do not jump-thread twice from the same block. */ | |
| 2437 if (bitmap_bit_p (threaded_blocks, entry->src->index)) | |
| 2438 { | |
| 2439 delete_jump_thread_path (path); | |
| 2440 paths.unordered_remove (i); | |
| 2441 } | |
| 2442 else | |
| 2443 i++; | |
| 2444 } | |
| 2445 | |
| 2446 bitmap_clear (threaded_blocks); | |
| 2447 | |
| 0 | 2448 mark_threaded_blocks (threaded_blocks); |
| 2449 | |
| 2450 initialize_original_copy_tables (); | |
| 2451 | |
| 2452 /* First perform the threading requests that do not affect | |
| 2453 loop structure. */ | |
| 2454 EXECUTE_IF_SET_IN_BITMAP (threaded_blocks, 0, i, bi) | |
| 2455 { | |
| 111 | 2456 basic_block bb = BASIC_BLOCK_FOR_FN (cfun, i); |
| 0 | 2457 |
| 2458 if (EDGE_COUNT (bb->preds) > 0) | |
| 2459 retval |= thread_block (bb, true); | |
| 2460 } | |
| 2461 | |
| 2462 /* Then perform the threading through loop headers. We start with the | |
| 2463 innermost loop, so that the changes in cfg we perform won't affect | |
| 2464 further threading. */ | |
| 111 | 2465 FOR_EACH_LOOP (loop, LI_FROM_INNERMOST) |
| 0 | 2466 { |
| 2467 if (!loop->header | |
| 2468 || !bitmap_bit_p (threaded_blocks, loop->header->index)) | |
| 2469 continue; | |
| 2470 | |
| 2471 retval |= thread_through_loop_header (loop, may_peel_loop_headers); | |
| 2472 } | |
| 2473 | |
| 111 | 2474 /* All jump threading paths should have been resolved at this |
| 2475 point. Verify that is the case. */ | |
| 2476 basic_block bb; | |
| 2477 FOR_EACH_BB_FN (bb, cfun) | |
| 2478 { | |
| 2479 edge_iterator ei; | |
| 2480 edge e; | |
| 2481 FOR_EACH_EDGE (e, ei, bb->preds) | |
| 2482 gcc_assert (e->aux == NULL); | |
| 2483 } | |
| 2484 | |
| 0 | 2485 statistics_counter_event (cfun, "Jumps threaded", |
| 2486 thread_stats.num_threaded_edges); | |
| 2487 | |
| 2488 free_original_copy_tables (); | |
| 2489 | |
| 111 | 2490 paths.release (); |
| 0 | 2491 |
| 2492 if (retval) | |
| 2493 loops_state_set (LOOPS_NEED_FIXUP); | |
| 2494 | |
| 111 | 2495 out: |
| 2496 delete removed_edges; | |
| 2497 removed_edges = NULL; | |
| 0 | 2498 return retval; |
| 2499 } | |
| 2500 | |
| 111 | 2501 /* Delete the jump threading path PATH. We have to explicitly delete |
| 2502 each entry in the vector, then the container. */ | |
| 2503 | |
| 2504 void | |
| 2505 delete_jump_thread_path (vec<jump_thread_edge *> *path) | |
| 2506 { | |
| 2507 for (unsigned int i = 0; i < path->length (); i++) | |
| 2508 delete (*path)[i]; | |
| 2509 path->release(); | |
| 2510 delete path; | |
| 2511 } | |
| 2512 | |
| 0 | 2513 /* Register a jump threading opportunity. We queue up all the jump |
| 2514 threading opportunities discovered by a pass and update the CFG | |
| 2515 and SSA form all at once. | |
| 2516 | |
| 2517 E is the edge we can thread, E2 is the new target edge, i.e., we | |
| 2518 are effectively recording that E->dest can be changed to E2->dest | |
| 2519 after fixing the SSA graph. */ | |
| 2520 | |
| 2521 void | |
| 111 | 2522 register_jump_thread (vec<jump_thread_edge *> *path) |
| 0 | 2523 { |
| 111 | 2524 if (!dbg_cnt (registered_jump_thread)) |
| 2525 { | |
| 2526 delete_jump_thread_path (path); | |
| 2527 return; | |
| 2528 } | |
| 2529 | |
| 2530 /* First make sure there are no NULL outgoing edges on the jump threading | |
| 2531 path. That can happen for jumping to a constant address. */ | |
| 2532 for (unsigned int i = 0; i < path->length (); i++) | |
| 2533 { | |
| 2534 if ((*path)[i]->e == NULL) | |
| 2535 { | |
| 2536 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 2537 { | |
| 2538 fprintf (dump_file, | |
| 2539 "Found NULL edge in jump threading path. Cancelling jump thread:\n"); | |
| 2540 dump_jump_thread_path (dump_file, *path, false); | |
| 2541 } | |
| 2542 | |
| 2543 delete_jump_thread_path (path); | |
| 2544 return; | |
| 2545 } | |
| 2546 | |
| 2547 /* Only the FSM threader is allowed to thread across | |
| 2548 backedges in the CFG. */ | |
| 2549 if (flag_checking | |
| 2550 && (*path)[0]->type != EDGE_FSM_THREAD) | |
| 2551 gcc_assert (((*path)[i]->e->flags & EDGE_DFS_BACK) == 0); | |
| 2552 } | |
| 2553 | |
| 2554 if (dump_file && (dump_flags & TDF_DETAILS)) | |
| 2555 dump_jump_thread_path (dump_file, *path, true); | |
| 2556 | |
| 2557 if (!paths.exists ()) | |
| 2558 paths.create (5); | |
| 2559 | |
| 2560 paths.safe_push (path); | |
| 0 | 2561 } |